License cleanup: add SPDX GPL-2.0 license identifier to files with no license
Many source files in the tree are missing licensing information, which
makes it harder for compliance tools to determine the correct license.
By default all files without license information are under the default
license of the kernel, which is GPL version 2.
Update the files which contain no license information with the 'GPL-2.0'
SPDX license identifier. The SPDX identifier is a legally binding
shorthand, which can be used instead of the full boiler plate text.
This patch is based on work done by Thomas Gleixner and Kate Stewart and
Philippe Ombredanne.
How this work was done:
Patches were generated and checked against linux-4.14-rc6 for a subset of
the use cases:
- file had no licensing information it it.
- file was a */uapi/* one with no licensing information in it,
- file was a */uapi/* one with existing licensing information,
Further patches will be generated in subsequent months to fix up cases
where non-standard license headers were used, and references to license
had to be inferred by heuristics based on keywords.
The analysis to determine which SPDX License Identifier to be applied to
a file was done in a spreadsheet of side by side results from of the
output of two independent scanners (ScanCode & Windriver) producing SPDX
tag:value files created by Philippe Ombredanne. Philippe prepared the
base worksheet, and did an initial spot review of a few 1000 files.
The 4.13 kernel was the starting point of the analysis with 60,537 files
assessed. Kate Stewart did a file by file comparison of the scanner
results in the spreadsheet to determine which SPDX license identifier(s)
to be applied to the file. She confirmed any determination that was not
immediately clear with lawyers working with the Linux Foundation.
Criteria used to select files for SPDX license identifier tagging was:
- Files considered eligible had to be source code files.
- Make and config files were included as candidates if they contained >5
lines of source
- File already had some variant of a license header in it (even if <5
lines).
All documentation files were explicitly excluded.
The following heuristics were used to determine which SPDX license
identifiers to apply.
- when both scanners couldn't find any license traces, file was
considered to have no license information in it, and the top level
COPYING file license applied.
For non */uapi/* files that summary was:
SPDX license identifier # files
---------------------------------------------------|-------
GPL-2.0 11139
and resulted in the first patch in this series.
If that file was a */uapi/* path one, it was "GPL-2.0 WITH
Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was:
SPDX license identifier # files
---------------------------------------------------|-------
GPL-2.0 WITH Linux-syscall-note 930
and resulted in the second patch in this series.
- if a file had some form of licensing information in it, and was one
of the */uapi/* ones, it was denoted with the Linux-syscall-note if
any GPL family license was found in the file or had no licensing in
it (per prior point). Results summary:
SPDX license identifier # files
---------------------------------------------------|------
GPL-2.0 WITH Linux-syscall-note 270
GPL-2.0+ WITH Linux-syscall-note 169
((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21
((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17
LGPL-2.1+ WITH Linux-syscall-note 15
GPL-1.0+ WITH Linux-syscall-note 14
((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5
LGPL-2.0+ WITH Linux-syscall-note 4
LGPL-2.1 WITH Linux-syscall-note 3
((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3
((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1
and that resulted in the third patch in this series.
- when the two scanners agreed on the detected license(s), that became
the concluded license(s).
- when there was disagreement between the two scanners (one detected a
license but the other didn't, or they both detected different
licenses) a manual inspection of the file occurred.
- In most cases a manual inspection of the information in the file
resulted in a clear resolution of the license that should apply (and
which scanner probably needed to revisit its heuristics).
- When it was not immediately clear, the license identifier was
confirmed with lawyers working with the Linux Foundation.
- If there was any question as to the appropriate license identifier,
the file was flagged for further research and to be revisited later
in time.
In total, over 70 hours of logged manual review was done on the
spreadsheet to determine the SPDX license identifiers to apply to the
source files by Kate, Philippe, Thomas and, in some cases, confirmation
by lawyers working with the Linux Foundation.
Kate also obtained a third independent scan of the 4.13 code base from
FOSSology, and compared selected files where the other two scanners
disagreed against that SPDX file, to see if there was new insights. The
Windriver scanner is based on an older version of FOSSology in part, so
they are related.
Thomas did random spot checks in about 500 files from the spreadsheets
for the uapi headers and agreed with SPDX license identifier in the
files he inspected. For the non-uapi files Thomas did random spot checks
in about 15000 files.
In initial set of patches against 4.14-rc6, 3 files were found to have
copy/paste license identifier errors, and have been fixed to reflect the
correct identifier.
Additionally Philippe spent 10 hours this week doing a detailed manual
inspection and review of the 12,461 patched files from the initial patch
version early this week with:
- a full scancode scan run, collecting the matched texts, detected
license ids and scores
- reviewing anything where there was a license detected (about 500+
files) to ensure that the applied SPDX license was correct
- reviewing anything where there was no detection but the patch license
was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied
SPDX license was correct
This produced a worksheet with 20 files needing minor correction. This
worksheet was then exported into 3 different .csv files for the
different types of files to be modified.
These .csv files were then reviewed by Greg. Thomas wrote a script to
parse the csv files and add the proper SPDX tag to the file, in the
format that the file expected. This script was further refined by Greg
based on the output to detect more types of files automatically and to
distinguish between header and source .c files (which need different
comment types.) Finally Greg ran the script using the .csv files to
generate the patches.
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 14:07:57 +00:00
|
|
|
// SPDX-License-Identifier: GPL-2.0
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
|
|
|
|
*
|
|
|
|
* Swap reorganised 29.12.95, Stephen Tweedie.
|
|
|
|
* kswapd added: 7.1.96 sct
|
|
|
|
* Removed kswapd_ctl limits, and swap out as many pages as needed
|
|
|
|
* to bring the system back to freepages.high: 2.4.97, Rik van Riel.
|
|
|
|
* Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
|
|
|
|
* Multiqueue VM started 5.8.00, Rik van Riel.
|
|
|
|
*/
|
|
|
|
|
2014-06-06 21:38:30 +00:00
|
|
|
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/mm.h>
|
2017-02-02 19:43:54 +00:00
|
|
|
#include <linux/sched/mm.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/module.h>
|
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 08:04:11 +00:00
|
|
|
#include <linux/gfp.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/kernel_stat.h>
|
|
|
|
#include <linux/swap.h>
|
|
|
|
#include <linux/pagemap.h>
|
|
|
|
#include <linux/init.h>
|
|
|
|
#include <linux/highmem.h>
|
memcg: add memory.pressure_level events
With this patch userland applications that want to maintain the
interactivity/memory allocation cost can use the pressure level
notifications. The levels are defined like this:
The "low" level means that the system is reclaiming memory for new
allocations. Monitoring this reclaiming activity might be useful for
maintaining cache level. Upon notification, the program (typically
"Activity Manager") might analyze vmstat and act in advance (i.e.
prematurely shutdown unimportant services).
The "medium" level means that the system is experiencing medium memory
pressure, the system might be making swap, paging out active file
caches, etc. Upon this event applications may decide to further analyze
vmstat/zoneinfo/memcg or internal memory usage statistics and free any
resources that can be easily reconstructed or re-read from a disk.
The "critical" level means that the system is actively thrashing, it is
about to out of memory (OOM) or even the in-kernel OOM killer is on its
way to trigger. Applications should do whatever they can to help the
system. It might be too late to consult with vmstat or any other
statistics, so it's advisable to take an immediate action.
The events are propagated upward until the event is handled, i.e. the
events are not pass-through. Here is what this means: for example you
have three cgroups: A->B->C. Now you set up an event listener on
cgroups A, B and C, and suppose group C experiences some pressure. In
this situation, only group C will receive the notification, i.e. groups
A and B will not receive it. This is done to avoid excessive
"broadcasting" of messages, which disturbs the system and which is
especially bad if we are low on memory or thrashing. So, organize the
cgroups wisely, or propagate the events manually (or, ask us to
implement the pass-through events, explaining why would you need them.)
Performance wise, the memory pressure notifications feature itself is
lightweight and does not require much of bookkeeping, in contrast to the
rest of memcg features. Unfortunately, as of current memcg
implementation, pages accounting is an inseparable part and cannot be
turned off. The good news is that there are some efforts[1] to improve
the situation; plus, implementing the same, fully API-compatible[2]
interface for CONFIG_MEMCG=n case (e.g. embedded) is also a viable
option, so it will not require any changes on the userland side.
[1] http://permalink.gmane.org/gmane.linux.kernel.cgroups/6291
[2] http://lkml.org/lkml/2013/2/21/454
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix CONFIG_CGROPUPS=n warnings]
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Leonid Moiseichuk <leonid.moiseichuk@nokia.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-29 22:08:31 +00:00
|
|
|
#include <linux/vmpressure.h>
|
2006-09-27 08:50:00 +00:00
|
|
|
#include <linux/vmstat.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/file.h>
|
|
|
|
#include <linux/writeback.h>
|
|
|
|
#include <linux/blkdev.h>
|
2022-06-17 15:42:47 +00:00
|
|
|
#include <linux/buffer_head.h> /* for buffer_heads_over_limit */
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/mm_inline.h>
|
|
|
|
#include <linux/backing-dev.h>
|
|
|
|
#include <linux/rmap.h>
|
|
|
|
#include <linux/topology.h>
|
|
|
|
#include <linux/cpu.h>
|
|
|
|
#include <linux/cpuset.h>
|
2011-01-13 23:45:56 +00:00
|
|
|
#include <linux/compaction.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/notifier.h>
|
|
|
|
#include <linux/rwsem.h>
|
2006-03-22 08:09:04 +00:00
|
|
|
#include <linux/delay.h>
|
2006-06-27 09:53:33 +00:00
|
|
|
#include <linux/kthread.h>
|
2006-12-07 04:34:23 +00:00
|
|
|
#include <linux/freezer.h>
|
2008-02-07 08:13:56 +00:00
|
|
|
#include <linux/memcontrol.h>
|
2021-09-02 21:59:16 +00:00
|
|
|
#include <linux/migrate.h>
|
2008-07-25 08:48:52 +00:00
|
|
|
#include <linux/delayacct.h>
|
2008-10-19 03:26:53 +00:00
|
|
|
#include <linux/sysctl.h>
|
2022-08-18 13:10:34 +00:00
|
|
|
#include <linux/memory-tiers.h>
|
vmscan: all_unreclaimable() use zone->all_unreclaimable as a name
all_unreclaimable check in direct reclaim has been introduced at 2.6.19
by following commit.
2006 Sep 25; commit 408d8544; oom: use unreclaimable info
And it went through strange history. firstly, following commit broke
the logic unintentionally.
2008 Apr 29; commit a41f24ea; page allocator: smarter retry of
costly-order allocations
Two years later, I've found obvious meaningless code fragment and
restored original intention by following commit.
2010 Jun 04; commit bb21c7ce; vmscan: fix do_try_to_free_pages()
return value when priority==0
But, the logic didn't works when 32bit highmem system goes hibernation
and Minchan slightly changed the algorithm and fixed it .
2010 Sep 22: commit d1908362: vmscan: check all_unreclaimable
in direct reclaim path
But, recently, Andrey Vagin found the new corner case. Look,
struct zone {
..
int all_unreclaimable;
..
unsigned long pages_scanned;
..
}
zone->all_unreclaimable and zone->pages_scanned are neigher atomic
variables nor protected by lock. Therefore zones can become a state of
zone->page_scanned=0 and zone->all_unreclaimable=1. In this case, current
all_unreclaimable() return false even though zone->all_unreclaimabe=1.
This resulted in the kernel hanging up when executing a loop of the form
1. fork
2. mmap
3. touch memory
4. read memory
5. munmmap
as described in
http://www.gossamer-threads.com/lists/linux/kernel/1348725#1348725
Is this ignorable minor issue? No. Unfortunately, x86 has very small dma
zone and it become zone->all_unreclamble=1 easily. and if it become
all_unreclaimable=1, it never restore all_unreclaimable=0. Why? if
all_unreclaimable=1, vmscan only try DEF_PRIORITY reclaim and
a-few-lru-pages>>DEF_PRIORITY always makes 0. that mean no page scan at
all!
Eventually, oom-killer never works on such systems. That said, we can't
use zone->pages_scanned for this purpose. This patch restore
all_unreclaimable() use zone->all_unreclaimable as old. and in addition,
to add oom_killer_disabled check to avoid reintroduce the issue of commit
d1908362 ("vmscan: check all_unreclaimable in direct reclaim path").
Reported-by: Andrey Vagin <avagin@openvz.org>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: David Rientjes <rientjes@google.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-04-14 22:22:12 +00:00
|
|
|
#include <linux/oom.h>
|
2018-11-06 13:23:24 +00:00
|
|
|
#include <linux/pagevec.h>
|
2011-05-20 19:50:29 +00:00
|
|
|
#include <linux/prefetch.h>
|
2014-06-06 21:38:30 +00:00
|
|
|
#include <linux/printk.h>
|
2016-01-22 23:10:40 +00:00
|
|
|
#include <linux/dax.h>
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:06:27 +00:00
|
|
|
#include <linux/psi.h>
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
#include <linux/pagewalk.h>
|
|
|
|
#include <linux/shmem_fs.h>
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
#include <linux/ctype.h>
|
2022-09-18 08:00:09 +00:00
|
|
|
#include <linux/debugfs.h>
|
2022-10-26 18:01:33 +00:00
|
|
|
#include <linux/khugepaged.h>
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
#include <linux/rculist_nulls.h>
|
|
|
|
#include <linux/random.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
#include <asm/tlbflush.h>
|
|
|
|
#include <asm/div64.h>
|
|
|
|
|
|
|
|
#include <linux/swapops.h>
|
2013-09-30 20:45:16 +00:00
|
|
|
#include <linux/balloon_compaction.h>
|
NUMA balancing: optimize page placement for memory tiering system
With the advent of various new memory types, some machines will have
multiple types of memory, e.g. DRAM and PMEM (persistent memory). The
memory subsystem of these machines can be called memory tiering system,
because the performance of the different types of memory are usually
different.
In such system, because of the memory accessing pattern changing etc,
some pages in the slow memory may become hot globally. So in this
patch, the NUMA balancing mechanism is enhanced to optimize the page
placement among the different memory types according to hot/cold
dynamically.
In a typical memory tiering system, there are CPUs, fast memory and slow
memory in each physical NUMA node. The CPUs and the fast memory will be
put in one logical node (called fast memory node), while the slow memory
will be put in another (faked) logical node (called slow memory node).
That is, the fast memory is regarded as local while the slow memory is
regarded as remote. So it's possible for the recently accessed pages in
the slow memory node to be promoted to the fast memory node via the
existing NUMA balancing mechanism.
The original NUMA balancing mechanism will stop to migrate pages if the
free memory of the target node becomes below the high watermark. This
is a reasonable policy if there's only one memory type. But this makes
the original NUMA balancing mechanism almost do not work to optimize
page placement among different memory types. Details are as follows.
It's the common cases that the working-set size of the workload is
larger than the size of the fast memory nodes. Otherwise, it's
unnecessary to use the slow memory at all. So, there are almost always
no enough free pages in the fast memory nodes, so that the globally hot
pages in the slow memory node cannot be promoted to the fast memory
node. To solve the issue, we have 2 choices as follows,
a. Ignore the free pages watermark checking when promoting hot pages
from the slow memory node to the fast memory node. This will
create some memory pressure in the fast memory node, thus trigger
the memory reclaiming. So that, the cold pages in the fast memory
node will be demoted to the slow memory node.
b. Define a new watermark called wmark_promo which is higher than
wmark_high, and have kswapd reclaiming pages until free pages reach
such watermark. The scenario is as follows: when we want to promote
hot-pages from a slow memory to a fast memory, but fast memory's free
pages would go lower than high watermark with such promotion, we wake
up kswapd with wmark_promo watermark in order to demote cold pages and
free us up some space. So, next time we want to promote hot-pages we
might have a chance of doing so.
The choice "a" may create high memory pressure in the fast memory node.
If the memory pressure of the workload is high, the memory pressure
may become so high that the memory allocation latency of the workload
is influenced, e.g. the direct reclaiming may be triggered.
The choice "b" works much better at this aspect. If the memory
pressure of the workload is high, the hot pages promotion will stop
earlier because its allocation watermark is higher than that of the
normal memory allocation. So in this patch, choice "b" is implemented.
A new zone watermark (WMARK_PROMO) is added. Which is larger than the
high watermark and can be controlled via watermark_scale_factor.
In addition to the original page placement optimization among sockets,
the NUMA balancing mechanism is extended to be used to optimize page
placement according to hot/cold among different memory types. So the
sysctl user space interface (numa_balancing) is extended in a backward
compatible way as follow, so that the users can enable/disable these
functionality individually.
The sysctl is converted from a Boolean value to a bits field. The
definition of the flags is,
- 0: NUMA_BALANCING_DISABLED
- 1: NUMA_BALANCING_NORMAL
- 2: NUMA_BALANCING_MEMORY_TIERING
We have tested the patch with the pmbench memory accessing benchmark
with the 80:20 read/write ratio and the Gauss access address
distribution on a 2 socket Intel server with Optane DC Persistent
Memory Model. The test results shows that the pmbench score can
improve up to 95.9%.
Thanks Andrew Morton to help fix the document format error.
Link: https://lkml.kernel.org/r/20220221084529.1052339-3-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: zhongjiang-ali <zhongjiang-ali@linux.alibaba.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Feng Tang <feng.tang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-22 21:46:23 +00:00
|
|
|
#include <linux/sched/sysctl.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2006-03-22 08:08:33 +00:00
|
|
|
#include "internal.h"
|
2022-05-10 01:20:47 +00:00
|
|
|
#include "swap.h"
|
2006-03-22 08:08:33 +00:00
|
|
|
|
2010-08-10 00:19:16 +00:00
|
|
|
#define CREATE_TRACE_POINTS
|
|
|
|
#include <trace/events/vmscan.h>
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
struct scan_control {
|
2009-12-15 01:59:10 +00:00
|
|
|
/* How many pages shrink_list() should reclaim */
|
|
|
|
unsigned long nr_to_reclaim;
|
|
|
|
|
2014-08-06 23:06:19 +00:00
|
|
|
/*
|
|
|
|
* Nodemask of nodes allowed by the caller. If NULL, all nodes
|
|
|
|
* are scanned.
|
|
|
|
*/
|
|
|
|
nodemask_t *nodemask;
|
2012-05-29 22:06:57 +00:00
|
|
|
|
2012-01-13 01:17:52 +00:00
|
|
|
/*
|
|
|
|
* The memory cgroup that hit its limit and as a result is the
|
|
|
|
* primary target of this reclaim invocation.
|
|
|
|
*/
|
|
|
|
struct mem_cgroup *target_mem_cgroup;
|
2008-02-07 08:13:56 +00:00
|
|
|
|
mm: vmscan: determine anon/file pressure balance at the reclaim root
We split the LRU lists into anon and file, and we rebalance the scan
pressure between them when one of them begins thrashing: if the file cache
experiences workingset refaults, we increase the pressure on anonymous
pages; if the workload is stalled on swapins, we increase the pressure on
the file cache instead.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, LRU pressure balancing is
done on an individual cgroup LRU level. As a result, when one cgroup is
thrashing on the filesystem cache while a sibling may have cold anonymous
pages, pressure doesn't get equalized between them.
This patch moves LRU balancing decision to the root of reclaim - the same
level where the LRU order is established.
It does this by tracking LRU cost recursively, so that every level of the
cgroup tree knows the aggregate LRU cost of all memory within its domain.
When the page scanner calculates the scan balance for any given individual
cgroup's LRU list, it uses the values from the ancestor cgroup that
initiated the reclaim cycle.
If one sibling is then thrashing on the cache, it will tip the pressure
balance inside its ancestors, and the next hierarchical reclaim iteration
will go more after the anon pages in the tree.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@surriel.com>
Link: http://lkml.kernel.org/r/20200520232525.798933-13-hannes@cmpxchg.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:03:06 +00:00
|
|
|
/*
|
|
|
|
* Scan pressure balancing between anon and file LRUs
|
|
|
|
*/
|
|
|
|
unsigned long anon_cost;
|
|
|
|
unsigned long file_cost;
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
/* Can active folios be deactivated as part of reclaim? */
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
#define DEACTIVATE_ANON 1
|
|
|
|
#define DEACTIVATE_FILE 2
|
|
|
|
unsigned int may_deactivate:2;
|
|
|
|
unsigned int force_deactivate:1;
|
|
|
|
unsigned int skipped_deactivate:1;
|
|
|
|
|
mm: vmscan: scan dirty pages even in laptop mode
Patch series "mm: vmscan: fix kswapd writeback regression".
We noticed a regression on multiple hadoop workloads when moving from
3.10 to 4.0 and 4.6, which involves kswapd getting tangled up in page
writeout, causing direct reclaim herds that also don't make progress.
I tracked it down to the thrash avoidance efforts after 3.10 that make
the kernel better at keeping use-once cache and use-many cache sorted on
the inactive and active list, with more aggressive protection of the
active list as long as there is inactive cache. Unfortunately, our
workload's use-once cache is mostly from streaming writes. Waiting for
writes to avoid potential reloads in the future is not a good tradeoff.
These patches do the following:
1. Wake the flushers when kswapd sees a lump of dirty pages. It's
possible to be below the dirty background limit and still have cache
velocity push them through the LRU. So start a-flushin'.
2. Let kswapd only write pages that have been rotated twice. This makes
sure we really tried to get all the clean pages on the inactive list
before resorting to horrible LRU-order writeback.
3. Move rotating dirty pages off the inactive list. Instead of churning
or waiting on page writeback, we'll go after clean active cache. This
might lead to thrashing, but in this state memory demand outstrips IO
speed anyway, and reads are faster than writes.
Mel backported the series to 4.10-rc5 with one minor conflict and ran a
couple of tests on it. Mix of read/write random workload didn't show
anything interesting. Write-only database didn't show much difference
in performance but there were slight reductions in IO -- probably in the
noise.
simoop did show big differences although not as big as Mel expected.
This is Chris Mason's workload that similate the VM activity of hadoop.
Mel won't go through the full details but over the samples measured
during an hour it reported
4.10.0-rc5 4.10.0-rc5
vanilla johannes-v1r1
Amean p50-Read 21346531.56 ( 0.00%) 21697513.24 ( -1.64%)
Amean p95-Read 24700518.40 ( 0.00%) 25743268.98 ( -4.22%)
Amean p99-Read 27959842.13 ( 0.00%) 28963271.11 ( -3.59%)
Amean p50-Write 1138.04 ( 0.00%) 989.82 ( 13.02%)
Amean p95-Write 1106643.48 ( 0.00%) 12104.00 ( 98.91%)
Amean p99-Write 1569213.22 ( 0.00%) 36343.38 ( 97.68%)
Amean p50-Allocation 85159.82 ( 0.00%) 79120.70 ( 7.09%)
Amean p95-Allocation 204222.58 ( 0.00%) 129018.43 ( 36.82%)
Amean p99-Allocation 278070.04 ( 0.00%) 183354.43 ( 34.06%)
Amean final-p50-Read 21266432.00 ( 0.00%) 21921792.00 ( -3.08%)
Amean final-p95-Read 24870912.00 ( 0.00%) 26116096.00 ( -5.01%)
Amean final-p99-Read 28147712.00 ( 0.00%) 29523968.00 ( -4.89%)
Amean final-p50-Write 1130.00 ( 0.00%) 977.00 ( 13.54%)
Amean final-p95-Write 1033216.00 ( 0.00%) 2980.00 ( 99.71%)
Amean final-p99-Write 1517568.00 ( 0.00%) 32672.00 ( 97.85%)
Amean final-p50-Allocation 86656.00 ( 0.00%) 78464.00 ( 9.45%)
Amean final-p95-Allocation 211712.00 ( 0.00%) 116608.00 ( 44.92%)
Amean final-p99-Allocation 287232.00 ( 0.00%) 168704.00 ( 41.27%)
The latencies are actually completely horrific in comparison to 4.4 (and
4.10-rc5 is worse than 4.9 according to historical data for reasons Mel
hasn't analysed yet).
Still, 95% of write latency (p95-write) is halved by the series and
allocation latency is way down. Direct reclaim activity is one fifth of
what it was according to vmstats. Kswapd activity is higher but this is
not necessarily surprising. Kswapd efficiency is unchanged at 99% (99%
of pages scanned were reclaimed) but direct reclaim efficiency went from
77% to 99%
In the vanilla kernel, 627MB of data was written back from reclaim
context. With the series, no data was written back. With or without
the patch, pages are being immediately reclaimed after writeback
completes. However, with the patch, only 1/8th of the pages are
reclaimed like this.
This patch (of 5):
We have an elaborate dirty/writeback throttling mechanism inside the
reclaim scanner, but for that to work the pages have to go through
shrink_page_list() and get counted for what they are. Otherwise, we
mess up the LRU order and don't match reclaim speed to writeback.
Especially during deactivation, there is never a reason to skip dirty
pages; nothing is even trying to write them out from there. Don't mess
up the LRU order for nothing, shuffle these pages along.
Link: http://lkml.kernel.org/r/20170123181641.23938-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-24 22:56:11 +00:00
|
|
|
/* Writepage batching in laptop mode; RECLAIM_WRITE */
|
2014-08-06 23:06:19 +00:00
|
|
|
unsigned int may_writepage:1;
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
/* Can mapped folios be reclaimed? */
|
2014-08-06 23:06:19 +00:00
|
|
|
unsigned int may_unmap:1;
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
/* Can folios be swapped as part of reclaim? */
|
2014-08-06 23:06:19 +00:00
|
|
|
unsigned int may_swap:1;
|
|
|
|
|
2022-07-14 06:49:18 +00:00
|
|
|
/* Proactive reclaim invoked by userspace through memory.reclaim */
|
|
|
|
unsigned int proactive:1;
|
|
|
|
|
mm/vmscan: more restrictive condition for retry in do_try_to_free_pages
By reviewing code, I find that when enter do_try_to_free_pages, the
may_thrash is always clear, and it will retry shrink zones to tap
cgroup's reserves memory by setting may_thrash when the former
shrink_zones reclaim nothing.
However, when memcg is disabled or on legacy hierarchy, or there do not
have any memcg protected by low limit, it should not do this useless
retry at all, for we do not have any cgroup's reserves memory to tap,
and we have already done hard work but made no progress, which as Michal
pointed out in former version, we are trying hard to control the retry
logical of page alloctor, and the current additional round of reclaim is
just lame.
Therefore, to avoid this unneeded retrying and make code more readable,
we remove the may_thrash field in scan_control, instead, introduce
memcg_low_reclaim and memcg_low_skipped, and only retry when
memcg_low_skipped, by setting memcg_low_reclaim.
[xieyisheng1@huawei.com: remove may_thrash field, introduce mem_cgroup_reclaim]
Link: http://lkml.kernel.org/r/1490191893-5923-1-git-send-email-ysxie@foxmail.com
Link: http://lkml.kernel.org/r/1490191893-5923-1-git-send-email-ysxie@foxmail.com
Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:57 +00:00
|
|
|
/*
|
2021-08-20 02:04:21 +00:00
|
|
|
* Cgroup memory below memory.low is protected as long as we
|
|
|
|
* don't threaten to OOM. If any cgroup is reclaimed at
|
|
|
|
* reduced force or passed over entirely due to its memory.low
|
|
|
|
* setting (memcg_low_skipped), and nothing is reclaimed as a
|
|
|
|
* result, then go back for one more cycle that reclaims the protected
|
|
|
|
* memory (memcg_low_reclaim) to avert OOM.
|
mm/vmscan: more restrictive condition for retry in do_try_to_free_pages
By reviewing code, I find that when enter do_try_to_free_pages, the
may_thrash is always clear, and it will retry shrink zones to tap
cgroup's reserves memory by setting may_thrash when the former
shrink_zones reclaim nothing.
However, when memcg is disabled or on legacy hierarchy, or there do not
have any memcg protected by low limit, it should not do this useless
retry at all, for we do not have any cgroup's reserves memory to tap,
and we have already done hard work but made no progress, which as Michal
pointed out in former version, we are trying hard to control the retry
logical of page alloctor, and the current additional round of reclaim is
just lame.
Therefore, to avoid this unneeded retrying and make code more readable,
we remove the may_thrash field in scan_control, instead, introduce
memcg_low_reclaim and memcg_low_skipped, and only retry when
memcg_low_skipped, by setting memcg_low_reclaim.
[xieyisheng1@huawei.com: remove may_thrash field, introduce mem_cgroup_reclaim]
Link: http://lkml.kernel.org/r/1490191893-5923-1-git-send-email-ysxie@foxmail.com
Link: http://lkml.kernel.org/r/1490191893-5923-1-git-send-email-ysxie@foxmail.com
Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:57 +00:00
|
|
|
*/
|
|
|
|
unsigned int memcg_low_reclaim:1;
|
|
|
|
unsigned int memcg_low_skipped:1;
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 23:26:06 +00:00
|
|
|
|
2014-08-06 23:06:19 +00:00
|
|
|
unsigned int hibernation_mode:1;
|
|
|
|
|
|
|
|
/* One of the zones is ready for compaction */
|
|
|
|
unsigned int compaction_ready:1;
|
|
|
|
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
/* There is easily reclaimable cold cache in the current node */
|
|
|
|
unsigned int cache_trim_mode:1;
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
/* The file folios on the current node are dangerously low */
|
mm: vmscan: move file exhaustion detection to the node level
Patch series "mm: fix page aging across multiple cgroups".
When applications are put into unconfigured cgroups for memory accounting
purposes, the cgrouping itself should not change the behavior of the page
reclaim code. We expect the VM to reclaim the coldest pages in the
system. But right now the VM can reclaim hot pages in one cgroup while
there is eligible cold cache in others.
This is because one part of the reclaim algorithm isn't truly cgroup
hierarchy aware: the inactive/active list balancing. That is the part
that is supposed to protect hot cache data from one-off streaming IO.
The recursive cgroup reclaim scheme will scan and rotate the physical LRU
lists of each eligible cgroup at the same rate in a round-robin fashion,
thereby establishing a relative order among the pages of all those
cgroups. However, the inactive/active balancing decisions are made
locally within each cgroup, so when a cgroup is running low on cold pages,
its hot pages will get reclaimed - even when sibling cgroups have plenty
of cold cache eligible in the same reclaim run.
For example:
[root@ham ~]# head -n1 /proc/meminfo
MemTotal: 1016336 kB
[root@ham ~]# ./reclaimtest2.sh
Establishing 50M active files in cgroup A...
Hot pages cached: 12800/12800 workingset-a
Linearly scanning through 18G of file data in cgroup B:
real 0m4.269s
user 0m0.051s
sys 0m4.182s
Hot pages cached: 134/12800 workingset-a
The streaming IO in B, which doesn't benefit from caching at all, pushes
out most of the workingset in A.
Solution
This series fixes the problem by elevating inactive/active balancing
decisions to the toplevel of the reclaim run. This is either a cgroup
that hit its limit, or straight-up global reclaim if there is physical
memory pressure. From there, it takes a recursive view of the cgroup
subtree to decide whether page deactivation is necessary.
In the test above, the VM will then recognize that cgroup B has plenty of
eligible cold cache, and that the hot pages in A can be spared:
[root@ham ~]# ./reclaimtest2.sh
Establishing 50M active files in cgroup A...
Hot pages cached: 12800/12800 workingset-a
Linearly scanning through 18G of file data in cgroup B:
real 0m4.244s
user 0m0.064s
sys 0m4.177s
Hot pages cached: 12800/12800 workingset-a
Implementation
Whether active pages can be deactivated or not is influenced by two
factors: the inactive list dropping below a minimum size relative to the
active list, and the occurence of refaults.
This patch series first moves refault detection to the reclaim root, then
enforces the minimum inactive size based on a recursive view of the cgroup
tree's LRUs.
History
Note that this actually never worked correctly in Linux cgroups. In the
past it worked for global reclaim and leaf limit reclaim only (we used to
have two physical LRU linkages per page), but it never worked for
intermediate limit reclaim over multiple leaf cgroups.
We're noticing this now because 1) we're putting everything into cgroups
for accounting, not just the things we want to control and 2) we're moving
away from leaf limits that invoke reclaim on individual cgroups, toward
large tree reclaim, triggered by high-level limits, or physical memory
pressure that is influenced by local protections such as memory.low and
memory.min instead.
This patch (of 3):
When file pages are lower than the watermark on a node, we try to force
scan anonymous pages to counter-act the balancing algorithms preference
for new file pages when they are likely thrashing. This is a node-level
decision, but it's currently made each time we look at an lruvec. This is
unnecessarily expensive and also a layering violation that makes the code
harder to understand.
Clean this up by making the check once per node and setting a flag in the
scan_control.
Link: http://lkml.kernel.org/r/20191107205334.158354-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:56 +00:00
|
|
|
unsigned int file_is_tiny:1;
|
|
|
|
|
2021-09-02 21:59:16 +00:00
|
|
|
/* Always discard instead of demoting to lower tier memory */
|
|
|
|
unsigned int no_demotion:1;
|
|
|
|
|
2018-08-17 22:45:19 +00:00
|
|
|
/* Allocation order */
|
|
|
|
s8 order;
|
|
|
|
|
|
|
|
/* Scan (total_size >> priority) pages at once */
|
|
|
|
s8 priority;
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
/* The highest zone to isolate folios for reclaim from */
|
2018-08-17 22:45:19 +00:00
|
|
|
s8 reclaim_idx;
|
|
|
|
|
|
|
|
/* This context's GFP mask */
|
|
|
|
gfp_t gfp_mask;
|
|
|
|
|
2014-08-06 23:06:19 +00:00
|
|
|
/* Incremented by the number of inactive pages that were scanned */
|
|
|
|
unsigned long nr_scanned;
|
|
|
|
|
|
|
|
/* Number of pages freed so far during a call to shrink_zones() */
|
|
|
|
unsigned long nr_reclaimed;
|
2018-04-10 23:27:59 +00:00
|
|
|
|
|
|
|
struct {
|
|
|
|
unsigned int dirty;
|
|
|
|
unsigned int unqueued_dirty;
|
|
|
|
unsigned int congested;
|
|
|
|
unsigned int writeback;
|
|
|
|
unsigned int immediate;
|
|
|
|
unsigned int file_taken;
|
|
|
|
unsigned int taken;
|
|
|
|
} nr;
|
2019-07-16 23:26:09 +00:00
|
|
|
|
|
|
|
/* for recording the reclaimed slab by now */
|
|
|
|
struct reclaim_state reclaim_state;
|
2005-04-16 22:20:36 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
#ifdef ARCH_HAS_PREFETCHW
|
2022-06-17 15:42:45 +00:00
|
|
|
#define prefetchw_prev_lru_folio(_folio, _base, _field) \
|
2005-04-16 22:20:36 +00:00
|
|
|
do { \
|
2022-06-17 15:42:45 +00:00
|
|
|
if ((_folio)->lru.prev != _base) { \
|
|
|
|
struct folio *prev; \
|
2005-04-16 22:20:36 +00:00
|
|
|
\
|
2022-06-17 15:42:45 +00:00
|
|
|
prev = lru_to_folio(&(_folio->lru)); \
|
2005-04-16 22:20:36 +00:00
|
|
|
prefetchw(&prev->_field); \
|
|
|
|
} \
|
|
|
|
} while (0)
|
|
|
|
#else
|
2022-06-17 15:42:45 +00:00
|
|
|
#define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
|
2005-04-16 22:20:36 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
/*
|
2020-06-03 23:02:37 +00:00
|
|
|
* From 0 .. 200. Higher means more swappy.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
|
|
|
int vm_swappiness = 60;
|
|
|
|
|
2019-09-23 22:38:12 +00:00
|
|
|
static void set_task_reclaim_state(struct task_struct *task,
|
|
|
|
struct reclaim_state *rs)
|
|
|
|
{
|
|
|
|
/* Check for an overwrite */
|
|
|
|
WARN_ON_ONCE(rs && task->reclaim_state);
|
|
|
|
|
|
|
|
/* Check for the nulling of an already-nulled member */
|
|
|
|
WARN_ON_ONCE(!rs && !task->reclaim_state);
|
|
|
|
|
|
|
|
task->reclaim_state = rs;
|
|
|
|
}
|
|
|
|
|
2022-06-01 03:22:23 +00:00
|
|
|
LIST_HEAD(shrinker_list);
|
|
|
|
DECLARE_RWSEM(shrinker_rwsem);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2019-09-23 22:38:12 +00:00
|
|
|
#ifdef CONFIG_MEMCG
|
2021-05-05 01:36:17 +00:00
|
|
|
static int shrinker_nr_max;
|
2021-05-05 01:36:11 +00:00
|
|
|
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
/* The shrinker_info is expanded in a batch of BITS_PER_LONG */
|
2021-05-05 01:36:17 +00:00
|
|
|
static inline int shrinker_map_size(int nr_items)
|
|
|
|
{
|
|
|
|
return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
|
|
|
|
}
|
2021-05-05 01:36:11 +00:00
|
|
|
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
static inline int shrinker_defer_size(int nr_items)
|
|
|
|
{
|
|
|
|
return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:26 +00:00
|
|
|
static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
|
|
|
|
int nid)
|
|
|
|
{
|
|
|
|
return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
|
|
|
|
lockdep_is_held(&shrinker_rwsem));
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:23 +00:00
|
|
|
static int expand_one_shrinker_info(struct mem_cgroup *memcg,
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
int map_size, int defer_size,
|
|
|
|
int old_map_size, int old_defer_size)
|
2021-05-05 01:36:11 +00:00
|
|
|
{
|
2021-05-05 01:36:23 +00:00
|
|
|
struct shrinker_info *new, *old;
|
2021-05-05 01:36:11 +00:00
|
|
|
struct mem_cgroup_per_node *pn;
|
|
|
|
int nid;
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
int size = map_size + defer_size;
|
2021-05-05 01:36:11 +00:00
|
|
|
|
|
|
|
for_each_node(nid) {
|
|
|
|
pn = memcg->nodeinfo[nid];
|
2021-05-05 01:36:26 +00:00
|
|
|
old = shrinker_info_protected(memcg, nid);
|
2021-05-05 01:36:11 +00:00
|
|
|
/* Not yet online memcg */
|
|
|
|
if (!old)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
|
|
|
|
if (!new)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
new->nr_deferred = (atomic_long_t *)(new + 1);
|
|
|
|
new->map = (void *)new->nr_deferred + defer_size;
|
|
|
|
|
|
|
|
/* map: set all old bits, clear all new bits */
|
|
|
|
memset(new->map, (int)0xff, old_map_size);
|
|
|
|
memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
|
|
|
|
/* nr_deferred: copy old values, clear all new values */
|
|
|
|
memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
|
|
|
|
memset((void *)new->nr_deferred + old_defer_size, 0,
|
|
|
|
defer_size - old_defer_size);
|
2021-05-05 01:36:11 +00:00
|
|
|
|
2021-05-05 01:36:23 +00:00
|
|
|
rcu_assign_pointer(pn->shrinker_info, new);
|
2021-05-05 01:36:20 +00:00
|
|
|
kvfree_rcu(old, rcu);
|
2021-05-05 01:36:11 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:23 +00:00
|
|
|
void free_shrinker_info(struct mem_cgroup *memcg)
|
2021-05-05 01:36:11 +00:00
|
|
|
{
|
|
|
|
struct mem_cgroup_per_node *pn;
|
2021-05-05 01:36:23 +00:00
|
|
|
struct shrinker_info *info;
|
2021-05-05 01:36:11 +00:00
|
|
|
int nid;
|
|
|
|
|
|
|
|
for_each_node(nid) {
|
|
|
|
pn = memcg->nodeinfo[nid];
|
2021-05-05 01:36:23 +00:00
|
|
|
info = rcu_dereference_protected(pn->shrinker_info, true);
|
|
|
|
kvfree(info);
|
|
|
|
rcu_assign_pointer(pn->shrinker_info, NULL);
|
2021-05-05 01:36:11 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:23 +00:00
|
|
|
int alloc_shrinker_info(struct mem_cgroup *memcg)
|
2021-05-05 01:36:11 +00:00
|
|
|
{
|
2021-05-05 01:36:23 +00:00
|
|
|
struct shrinker_info *info;
|
2021-05-05 01:36:11 +00:00
|
|
|
int nid, size, ret = 0;
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
int map_size, defer_size = 0;
|
2021-05-05 01:36:11 +00:00
|
|
|
|
2021-05-05 01:36:14 +00:00
|
|
|
down_write(&shrinker_rwsem);
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
map_size = shrinker_map_size(shrinker_nr_max);
|
|
|
|
defer_size = shrinker_defer_size(shrinker_nr_max);
|
|
|
|
size = map_size + defer_size;
|
2021-05-05 01:36:11 +00:00
|
|
|
for_each_node(nid) {
|
2021-05-05 01:36:23 +00:00
|
|
|
info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
|
|
|
|
if (!info) {
|
|
|
|
free_shrinker_info(memcg);
|
2021-05-05 01:36:11 +00:00
|
|
|
ret = -ENOMEM;
|
|
|
|
break;
|
|
|
|
}
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
info->nr_deferred = (atomic_long_t *)(info + 1);
|
|
|
|
info->map = (void *)info->nr_deferred + defer_size;
|
2021-05-05 01:36:23 +00:00
|
|
|
rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
|
2021-05-05 01:36:11 +00:00
|
|
|
}
|
2021-05-05 01:36:14 +00:00
|
|
|
up_write(&shrinker_rwsem);
|
2021-05-05 01:36:11 +00:00
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
static inline bool need_expand(int nr_max)
|
|
|
|
{
|
|
|
|
return round_up(nr_max, BITS_PER_LONG) >
|
|
|
|
round_up(shrinker_nr_max, BITS_PER_LONG);
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:23 +00:00
|
|
|
static int expand_shrinker_info(int new_id)
|
2021-05-05 01:36:11 +00:00
|
|
|
{
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
int ret = 0;
|
2021-05-05 01:36:17 +00:00
|
|
|
int new_nr_max = new_id + 1;
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
int map_size, defer_size = 0;
|
|
|
|
int old_map_size, old_defer_size = 0;
|
2021-05-05 01:36:11 +00:00
|
|
|
struct mem_cgroup *memcg;
|
|
|
|
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
if (!need_expand(new_nr_max))
|
2021-05-05 01:36:17 +00:00
|
|
|
goto out;
|
2021-05-05 01:36:11 +00:00
|
|
|
|
|
|
|
if (!root_mem_cgroup)
|
2021-05-05 01:36:14 +00:00
|
|
|
goto out;
|
|
|
|
|
|
|
|
lockdep_assert_held(&shrinker_rwsem);
|
2021-05-05 01:36:11 +00:00
|
|
|
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
map_size = shrinker_map_size(new_nr_max);
|
|
|
|
defer_size = shrinker_defer_size(new_nr_max);
|
|
|
|
old_map_size = shrinker_map_size(shrinker_nr_max);
|
|
|
|
old_defer_size = shrinker_defer_size(shrinker_nr_max);
|
|
|
|
|
2021-05-05 01:36:11 +00:00
|
|
|
memcg = mem_cgroup_iter(NULL, NULL, NULL);
|
|
|
|
do {
|
mm: vmscan: add per memcg shrinker nr_deferred
Currently the number of deferred objects are per shrinker, but some
slabs, for example, vfs inode/dentry cache are per memcg, this would
result in poor isolation among memcgs.
The deferred objects typically are generated by __GFP_NOFS allocations,
one memcg with excessive __GFP_NOFS allocations may blow up deferred
objects, then other innocent memcgs may suffer from over shrink,
excessive reclaim latency, etc.
For example, two workloads run in memcgA and memcgB respectively,
workload in B is vfs heavy workload. Workload in A generates excessive
deferred objects, then B's vfs cache might be hit heavily (drop half of
caches) by B's limit reclaim or global reclaim.
We observed this hit in our production environment which was running vfs
heavy workload shown as the below tracing log:
<...>-409454 [016] .... 28286961.747146: mm_shrink_slab_start: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 objects to shrink 3641681686040 gfp_flags GFP_HIGHUSER_MOVABLE|__GFP_ZERO pgs_scanned 1 lru_pgs 15721
cache items 246404277 delta 31345 total_scan 123202138
<...>-409454 [022] .... 28287105.928018: mm_shrink_slab_end: super_cache_scan+0x0/0x1a0 ffff9a83046f3458:
nid: 1 unused scan count 3641681686040 new scan count 3641798379189 total_scan 602
last shrinker return val 123186855
The vfs cache and page cache ratio was 10:1 on this machine, and half of
caches were dropped. This also resulted in significant amount of page
caches were dropped due to inodes eviction.
Make nr_deferred per memcg for memcg aware shrinkers would solve the
unfairness and bring better isolation.
The following patch will add nr_deferred to parent memcg when memcg
offline. To preserve nr_deferred when reparenting memcgs to root, root
memcg needs shrinker_info allocated too.
When memcg is not enabled (!CONFIG_MEMCG or memcg disabled), the
shrinker's nr_deferred would be used. And non memcg aware shrinkers use
shrinker's nr_deferred all the time.
Link: https://lkml.kernel.org/r/20210311190845.9708-10-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <guro@fb.com>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:33 +00:00
|
|
|
ret = expand_one_shrinker_info(memcg, map_size, defer_size,
|
|
|
|
old_map_size, old_defer_size);
|
2021-05-05 01:36:11 +00:00
|
|
|
if (ret) {
|
|
|
|
mem_cgroup_iter_break(NULL, memcg);
|
2021-05-05 01:36:14 +00:00
|
|
|
goto out;
|
2021-05-05 01:36:11 +00:00
|
|
|
}
|
|
|
|
} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
|
2021-05-05 01:36:14 +00:00
|
|
|
out:
|
2021-05-05 01:36:11 +00:00
|
|
|
if (!ret)
|
2021-05-05 01:36:17 +00:00
|
|
|
shrinker_nr_max = new_nr_max;
|
2021-05-05 01:36:14 +00:00
|
|
|
|
2021-05-05 01:36:11 +00:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
|
|
|
|
{
|
|
|
|
if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
|
2021-05-05 01:36:23 +00:00
|
|
|
struct shrinker_info *info;
|
2021-05-05 01:36:11 +00:00
|
|
|
|
|
|
|
rcu_read_lock();
|
2021-05-05 01:36:23 +00:00
|
|
|
info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
|
2021-05-05 01:36:11 +00:00
|
|
|
/* Pairs with smp mb in shrink_slab() */
|
|
|
|
smp_mb__before_atomic();
|
2021-05-05 01:36:23 +00:00
|
|
|
set_bit(shrinker_id, info->map);
|
2021-05-05 01:36:11 +00:00
|
|
|
rcu_read_unlock();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-08-17 22:47:29 +00:00
|
|
|
static DEFINE_IDR(shrinker_idr);
|
|
|
|
|
|
|
|
static int prealloc_memcg_shrinker(struct shrinker *shrinker)
|
|
|
|
{
|
|
|
|
int id, ret = -ENOMEM;
|
|
|
|
|
2021-05-05 01:36:39 +00:00
|
|
|
if (mem_cgroup_disabled())
|
|
|
|
return -ENOSYS;
|
|
|
|
|
2018-08-17 22:47:29 +00:00
|
|
|
down_write(&shrinker_rwsem);
|
|
|
|
/* This may call shrinker, so it must use down_read_trylock() */
|
2021-05-05 01:36:29 +00:00
|
|
|
id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
|
2018-08-17 22:47:29 +00:00
|
|
|
if (id < 0)
|
|
|
|
goto unlock;
|
|
|
|
|
mm, memcg: assign memcg-aware shrinkers bitmap to memcg
Imagine a big node with many cpus, memory cgroups and containers. Let
we have 200 containers, every container has 10 mounts, and 10 cgroups.
All container tasks don't touch foreign containers mounts. If there is
intensive pages write, and global reclaim happens, a writing task has to
iterate over all memcgs to shrink slab, before it's able to go to
shrink_page_list().
Iteration over all the memcg slabs is very expensive: the task has to
visit 200 * 10 = 2000 shrinkers for every memcg, and since there are
2000 memcgs, the total calls are 2000 * 2000 = 4000000.
So, the shrinker makes 4 million do_shrink_slab() calls just to try to
isolate SWAP_CLUSTER_MAX pages in one of the actively writing memcg via
shrink_page_list(). I've observed a node spending almost 100% in
kernel, making useless iteration over already shrinked slab.
This patch adds bitmap of memcg-aware shrinkers to memcg. The size of
the bitmap depends on bitmap_nr_ids, and during memcg life it's
maintained to be enough to fit bitmap_nr_ids shrinkers. Every bit in
the map is related to corresponding shrinker id.
Next patches will maintain set bit only for really charged memcg. This
will allow shrink_slab() to increase its performance in significant way.
See the last patch for the numbers.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112549031.4097.3576147070498769979.stgit@localhost.localdomain
[ktkhai@virtuozzo.com: add comment to mem_cgroup_css_online()]
Link: http://lkml.kernel.org/r/521f9e5f-c436-b388-fe83-4dc870bfb489@virtuozzo.com
Link: http://lkml.kernel.org/r/153063056619.1818.12550500883688681076.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:47:37 +00:00
|
|
|
if (id >= shrinker_nr_max) {
|
2021-05-05 01:36:23 +00:00
|
|
|
if (expand_shrinker_info(id)) {
|
mm, memcg: assign memcg-aware shrinkers bitmap to memcg
Imagine a big node with many cpus, memory cgroups and containers. Let
we have 200 containers, every container has 10 mounts, and 10 cgroups.
All container tasks don't touch foreign containers mounts. If there is
intensive pages write, and global reclaim happens, a writing task has to
iterate over all memcgs to shrink slab, before it's able to go to
shrink_page_list().
Iteration over all the memcg slabs is very expensive: the task has to
visit 200 * 10 = 2000 shrinkers for every memcg, and since there are
2000 memcgs, the total calls are 2000 * 2000 = 4000000.
So, the shrinker makes 4 million do_shrink_slab() calls just to try to
isolate SWAP_CLUSTER_MAX pages in one of the actively writing memcg via
shrink_page_list(). I've observed a node spending almost 100% in
kernel, making useless iteration over already shrinked slab.
This patch adds bitmap of memcg-aware shrinkers to memcg. The size of
the bitmap depends on bitmap_nr_ids, and during memcg life it's
maintained to be enough to fit bitmap_nr_ids shrinkers. Every bit in
the map is related to corresponding shrinker id.
Next patches will maintain set bit only for really charged memcg. This
will allow shrink_slab() to increase its performance in significant way.
See the last patch for the numbers.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112549031.4097.3576147070498769979.stgit@localhost.localdomain
[ktkhai@virtuozzo.com: add comment to mem_cgroup_css_online()]
Link: http://lkml.kernel.org/r/521f9e5f-c436-b388-fe83-4dc870bfb489@virtuozzo.com
Link: http://lkml.kernel.org/r/153063056619.1818.12550500883688681076.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:47:37 +00:00
|
|
|
idr_remove(&shrinker_idr, id);
|
|
|
|
goto unlock;
|
|
|
|
}
|
|
|
|
}
|
2018-08-17 22:47:29 +00:00
|
|
|
shrinker->id = id;
|
|
|
|
ret = 0;
|
|
|
|
unlock:
|
|
|
|
up_write(&shrinker_rwsem);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void unregister_memcg_shrinker(struct shrinker *shrinker)
|
|
|
|
{
|
|
|
|
int id = shrinker->id;
|
|
|
|
|
|
|
|
BUG_ON(id < 0);
|
|
|
|
|
2021-05-05 01:36:29 +00:00
|
|
|
lockdep_assert_held(&shrinker_rwsem);
|
|
|
|
|
2018-08-17 22:47:29 +00:00
|
|
|
idr_remove(&shrinker_idr, id);
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:36 +00:00
|
|
|
static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
|
|
|
|
struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
struct shrinker_info *info;
|
|
|
|
|
|
|
|
info = shrinker_info_protected(memcg, nid);
|
|
|
|
return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
|
|
|
|
struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
struct shrinker_info *info;
|
|
|
|
|
|
|
|
info = shrinker_info_protected(memcg, nid);
|
|
|
|
return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:42 +00:00
|
|
|
void reparent_shrinker_deferred(struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
int i, nid;
|
|
|
|
long nr;
|
|
|
|
struct mem_cgroup *parent;
|
|
|
|
struct shrinker_info *child_info, *parent_info;
|
|
|
|
|
|
|
|
parent = parent_mem_cgroup(memcg);
|
|
|
|
if (!parent)
|
|
|
|
parent = root_mem_cgroup;
|
|
|
|
|
|
|
|
/* Prevent from concurrent shrinker_info expand */
|
|
|
|
down_read(&shrinker_rwsem);
|
|
|
|
for_each_node(nid) {
|
|
|
|
child_info = shrinker_info_protected(memcg, nid);
|
|
|
|
parent_info = shrinker_info_protected(parent, nid);
|
|
|
|
for (i = 0; i < shrinker_nr_max; i++) {
|
|
|
|
nr = atomic_long_read(&child_info->nr_deferred[i]);
|
|
|
|
atomic_long_add(nr, &parent_info->nr_deferred[i]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
up_read(&shrinker_rwsem);
|
|
|
|
}
|
|
|
|
|
2019-12-01 01:55:40 +00:00
|
|
|
static bool cgroup_reclaim(struct scan_control *sc)
|
2012-01-13 01:17:50 +00:00
|
|
|
{
|
2019-12-01 01:55:40 +00:00
|
|
|
return sc->target_mem_cgroup;
|
2012-01-13 01:17:50 +00:00
|
|
|
}
|
2015-05-22 22:23:36 +00:00
|
|
|
|
2022-12-22 04:19:01 +00:00
|
|
|
static bool global_reclaim(struct scan_control *sc)
|
|
|
|
{
|
|
|
|
return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup);
|
|
|
|
}
|
|
|
|
|
2015-05-22 22:23:36 +00:00
|
|
|
/**
|
2019-12-01 01:55:40 +00:00
|
|
|
* writeback_throttling_sane - is the usual dirty throttling mechanism available?
|
2015-05-22 22:23:36 +00:00
|
|
|
* @sc: scan_control in question
|
|
|
|
*
|
|
|
|
* The normal page dirty throttling mechanism in balance_dirty_pages() is
|
|
|
|
* completely broken with the legacy memcg and direct stalling in
|
2022-09-02 19:45:57 +00:00
|
|
|
* shrink_folio_list() is used for throttling instead, which lacks all the
|
2015-05-22 22:23:36 +00:00
|
|
|
* niceties such as fairness, adaptive pausing, bandwidth proportional
|
|
|
|
* allocation and configurability.
|
|
|
|
*
|
|
|
|
* This function tests whether the vmscan currently in progress can assume
|
|
|
|
* that the normal dirty throttling mechanism is operational.
|
|
|
|
*/
|
2019-12-01 01:55:40 +00:00
|
|
|
static bool writeback_throttling_sane(struct scan_control *sc)
|
2015-05-22 22:23:36 +00:00
|
|
|
{
|
2019-12-01 01:55:40 +00:00
|
|
|
if (!cgroup_reclaim(sc))
|
2015-05-22 22:23:36 +00:00
|
|
|
return true;
|
|
|
|
#ifdef CONFIG_CGROUP_WRITEBACK
|
2015-11-05 22:51:32 +00:00
|
|
|
if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
|
2015-05-22 22:23:36 +00:00
|
|
|
return true;
|
|
|
|
#endif
|
|
|
|
return false;
|
|
|
|
}
|
2008-02-07 08:14:29 +00:00
|
|
|
#else
|
2019-09-23 22:38:12 +00:00
|
|
|
static int prealloc_memcg_shrinker(struct shrinker *shrinker)
|
|
|
|
{
|
2021-05-05 01:36:39 +00:00
|
|
|
return -ENOSYS;
|
2019-09-23 22:38:12 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void unregister_memcg_shrinker(struct shrinker *shrinker)
|
|
|
|
{
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:36 +00:00
|
|
|
static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
|
|
|
|
struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
|
|
|
|
struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2019-12-01 01:55:40 +00:00
|
|
|
static bool cgroup_reclaim(struct scan_control *sc)
|
2012-01-13 01:17:50 +00:00
|
|
|
{
|
2019-12-01 01:55:40 +00:00
|
|
|
return false;
|
2012-01-13 01:17:50 +00:00
|
|
|
}
|
2015-05-22 22:23:36 +00:00
|
|
|
|
2022-12-22 04:19:01 +00:00
|
|
|
static bool global_reclaim(struct scan_control *sc)
|
|
|
|
{
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2019-12-01 01:55:40 +00:00
|
|
|
static bool writeback_throttling_sane(struct scan_control *sc)
|
2015-05-22 22:23:36 +00:00
|
|
|
{
|
|
|
|
return true;
|
|
|
|
}
|
2008-02-07 08:14:29 +00:00
|
|
|
#endif
|
|
|
|
|
2021-05-05 01:36:36 +00:00
|
|
|
static long xchg_nr_deferred(struct shrinker *shrinker,
|
|
|
|
struct shrink_control *sc)
|
|
|
|
{
|
|
|
|
int nid = sc->nid;
|
|
|
|
|
|
|
|
if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
|
|
|
|
nid = 0;
|
|
|
|
|
|
|
|
if (sc->memcg &&
|
|
|
|
(shrinker->flags & SHRINKER_MEMCG_AWARE))
|
|
|
|
return xchg_nr_deferred_memcg(nid, shrinker,
|
|
|
|
sc->memcg);
|
|
|
|
|
|
|
|
return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static long add_nr_deferred(long nr, struct shrinker *shrinker,
|
|
|
|
struct shrink_control *sc)
|
|
|
|
{
|
|
|
|
int nid = sc->nid;
|
|
|
|
|
|
|
|
if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
|
|
|
|
nid = 0;
|
|
|
|
|
|
|
|
if (sc->memcg &&
|
|
|
|
(shrinker->flags & SHRINKER_MEMCG_AWARE))
|
|
|
|
return add_nr_deferred_memcg(nr, nid, shrinker,
|
|
|
|
sc->memcg);
|
|
|
|
|
|
|
|
return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
|
|
|
|
}
|
|
|
|
|
2021-09-02 21:59:16 +00:00
|
|
|
static bool can_demote(int nid, struct scan_control *sc)
|
|
|
|
{
|
2021-09-02 21:59:33 +00:00
|
|
|
if (!numa_demotion_enabled)
|
|
|
|
return false;
|
Revert "mm/vmscan: never demote for memcg reclaim"
This reverts commit 3a235693d3930e1276c8d9cc0ca5807ef292cf0a.
Its premise was that cgroup reclaim cares about freeing memory inside the
cgroup, and demotion just moves them around within the cgroup limit.
Hence, pages from toptier nodes should be reclaimed directly.
However, with NUMA balancing now doing tier promotions, demotion is part
of the page aging process. Global reclaim demotes the coldest toptier
pages to secondary memory, where their life continues and from which they
have a chance to get promoted back. Essentially, tiered memory systems
have an LRU order that spans multiple nodes.
When cgroup reclaims pages coming off the toptier directly, there can be
colder pages on lower tier nodes that were demoted by global reclaim.
This is an aging inversion, not unlike if cgroups were to reclaim directly
from the active lists while there are inactive pages.
Proactive reclaim is another factor. The goal of that it is to offload
colder pages from expensive RAM to cheaper storage. When lower tier
memory is available as an intermediate layer, we want offloading to take
advantage of it instead of bypassing to storage.
Revert the patch so that cgroups respect the LRU order spanning the memory
hierarchy.
Of note is a specific undercommit scenario, where all cgroup limits in the
system add up to <= available toptier memory. In that case, shuffling
pages out to lower tiers first to reclaim them from there is inefficient.
This is something could be optimized/short-circuited later on (although
care must be taken not to accidentally recreate the aging inversion).
Let's ensure correctness first.
Link: https://lkml.kernel.org/r/20220518190911.82400-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Dave Hansen <dave.hansen@linux.intel.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Acked-by: Roman Gushchin <roman.gushchin@linux.dev>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Muchun Song <songmuchun@bytedance.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-18 19:09:11 +00:00
|
|
|
if (sc && sc->no_demotion)
|
|
|
|
return false;
|
2021-09-02 21:59:16 +00:00
|
|
|
if (next_demotion_node(nid) == NUMA_NO_NODE)
|
|
|
|
return false;
|
|
|
|
|
2021-09-02 21:59:33 +00:00
|
|
|
return true;
|
2021-09-02 21:59:16 +00:00
|
|
|
}
|
|
|
|
|
2021-09-02 21:59:26 +00:00
|
|
|
static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
|
|
|
|
int nid,
|
|
|
|
struct scan_control *sc)
|
|
|
|
{
|
|
|
|
if (memcg == NULL) {
|
|
|
|
/*
|
|
|
|
* For non-memcg reclaim, is there
|
|
|
|
* space in any swap device?
|
|
|
|
*/
|
|
|
|
if (get_nr_swap_pages() > 0)
|
|
|
|
return true;
|
|
|
|
} else {
|
|
|
|
/* Is the memcg below its swap limit? */
|
|
|
|
if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The page can not be swapped.
|
|
|
|
*
|
|
|
|
* Can it be reclaimed from this node via demotion?
|
|
|
|
*/
|
|
|
|
return can_demote(nid, sc);
|
|
|
|
}
|
|
|
|
|
2016-07-28 22:47:31 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* This misses isolated folios which are not accounted for to save counters.
|
2016-07-28 22:47:31 +00:00
|
|
|
* As the data only determines if reclaim or compaction continues, it is
|
2022-09-02 19:45:57 +00:00
|
|
|
* not expected that isolated folios will be a dominating factor.
|
2016-07-28 22:47:31 +00:00
|
|
|
*/
|
|
|
|
unsigned long zone_reclaimable_pages(struct zone *zone)
|
|
|
|
{
|
|
|
|
unsigned long nr;
|
|
|
|
|
|
|
|
nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
|
|
|
|
zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
|
2021-09-02 21:59:26 +00:00
|
|
|
if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
|
2016-07-28 22:47:31 +00:00
|
|
|
nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
|
|
|
|
zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
|
|
|
|
|
|
|
|
return nr;
|
|
|
|
}
|
|
|
|
|
2017-02-22 23:45:58 +00:00
|
|
|
/**
|
|
|
|
* lruvec_lru_size - Returns the number of pages on the given LRU list.
|
|
|
|
* @lruvec: lru vector
|
|
|
|
* @lru: lru to use
|
2022-04-29 06:16:04 +00:00
|
|
|
* @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
|
2017-02-22 23:45:58 +00:00
|
|
|
*/
|
2021-02-24 20:08:44 +00:00
|
|
|
static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
|
|
|
|
int zone_idx)
|
2009-01-08 02:08:16 +00:00
|
|
|
{
|
2019-12-01 01:55:31 +00:00
|
|
|
unsigned long size = 0;
|
2017-02-22 23:45:58 +00:00
|
|
|
int zid;
|
|
|
|
|
2022-04-29 06:16:04 +00:00
|
|
|
for (zid = 0; zid <= zone_idx; zid++) {
|
2017-02-22 23:45:58 +00:00
|
|
|
struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
|
2009-01-08 02:08:16 +00:00
|
|
|
|
2017-02-22 23:45:58 +00:00
|
|
|
if (!managed_zone(zone))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (!mem_cgroup_disabled())
|
2019-12-01 01:55:31 +00:00
|
|
|
size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
|
2017-02-22 23:45:58 +00:00
|
|
|
else
|
2019-12-01 01:55:31 +00:00
|
|
|
size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
|
2017-02-22 23:45:58 +00:00
|
|
|
}
|
2019-12-01 01:55:31 +00:00
|
|
|
return size;
|
2017-01-11 00:58:04 +00:00
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
* Add a shrinker callback to be called from the vm.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2022-06-01 03:22:24 +00:00
|
|
|
static int __prealloc_shrinker(struct shrinker *shrinker)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2021-05-05 01:36:39 +00:00
|
|
|
unsigned int size;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
|
|
|
|
err = prealloc_memcg_shrinker(shrinker);
|
|
|
|
if (err != -ENOSYS)
|
|
|
|
return err;
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
2021-05-05 01:36:39 +00:00
|
|
|
shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
|
|
|
|
}
|
|
|
|
|
|
|
|
size = sizeof(*shrinker->nr_deferred);
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
if (shrinker->flags & SHRINKER_NUMA_AWARE)
|
|
|
|
size *= nr_node_ids;
|
|
|
|
|
|
|
|
shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
|
|
|
|
if (!shrinker->nr_deferred)
|
|
|
|
return -ENOMEM;
|
2018-08-17 22:47:29 +00:00
|
|
|
|
2018-04-04 10:53:07 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2022-06-01 03:22:24 +00:00
|
|
|
#ifdef CONFIG_SHRINKER_DEBUG
|
|
|
|
int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
|
|
|
|
{
|
|
|
|
va_list ap;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
va_start(ap, fmt);
|
|
|
|
shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
|
|
|
|
va_end(ap);
|
|
|
|
if (!shrinker->name)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
err = __prealloc_shrinker(shrinker);
|
2022-07-20 14:47:55 +00:00
|
|
|
if (err) {
|
2022-06-01 03:22:24 +00:00
|
|
|
kfree_const(shrinker->name);
|
2022-07-20 14:47:55 +00:00
|
|
|
shrinker->name = NULL;
|
|
|
|
}
|
2022-06-01 03:22:24 +00:00
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
|
|
|
|
{
|
|
|
|
return __prealloc_shrinker(shrinker);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2018-04-04 10:53:07 +00:00
|
|
|
void free_prealloced_shrinker(struct shrinker *shrinker)
|
|
|
|
{
|
2022-06-01 03:22:24 +00:00
|
|
|
#ifdef CONFIG_SHRINKER_DEBUG
|
|
|
|
kfree_const(shrinker->name);
|
2022-07-20 14:47:55 +00:00
|
|
|
shrinker->name = NULL;
|
2022-06-01 03:22:24 +00:00
|
|
|
#endif
|
2021-05-05 01:36:29 +00:00
|
|
|
if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
|
|
|
|
down_write(&shrinker_rwsem);
|
2018-08-17 22:47:29 +00:00
|
|
|
unregister_memcg_shrinker(shrinker);
|
2021-05-05 01:36:29 +00:00
|
|
|
up_write(&shrinker_rwsem);
|
2021-05-05 01:36:39 +00:00
|
|
|
return;
|
2021-05-05 01:36:29 +00:00
|
|
|
}
|
2018-08-17 22:47:29 +00:00
|
|
|
|
2018-04-04 10:53:07 +00:00
|
|
|
kfree(shrinker->nr_deferred);
|
|
|
|
shrinker->nr_deferred = NULL;
|
|
|
|
}
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
2018-04-04 10:53:07 +00:00
|
|
|
void register_shrinker_prepared(struct shrinker *shrinker)
|
|
|
|
{
|
2007-07-17 11:03:17 +00:00
|
|
|
down_write(&shrinker_rwsem);
|
|
|
|
list_add_tail(&shrinker->list, &shrinker_list);
|
2021-05-05 01:36:29 +00:00
|
|
|
shrinker->flags |= SHRINKER_REGISTERED;
|
2022-06-01 03:22:23 +00:00
|
|
|
shrinker_debugfs_add(shrinker);
|
2007-07-17 11:03:17 +00:00
|
|
|
up_write(&shrinker_rwsem);
|
2018-04-04 10:53:07 +00:00
|
|
|
}
|
|
|
|
|
2022-06-01 03:22:24 +00:00
|
|
|
static int __register_shrinker(struct shrinker *shrinker)
|
2018-04-04 10:53:07 +00:00
|
|
|
{
|
2022-06-01 03:22:24 +00:00
|
|
|
int err = __prealloc_shrinker(shrinker);
|
2018-04-04 10:53:07 +00:00
|
|
|
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
register_shrinker_prepared(shrinker);
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
return 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2022-06-01 03:22:24 +00:00
|
|
|
|
|
|
|
#ifdef CONFIG_SHRINKER_DEBUG
|
|
|
|
int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
|
|
|
|
{
|
|
|
|
va_list ap;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
va_start(ap, fmt);
|
|
|
|
shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
|
|
|
|
va_end(ap);
|
|
|
|
if (!shrinker->name)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
err = __register_shrinker(shrinker);
|
2022-07-20 14:47:55 +00:00
|
|
|
if (err) {
|
2022-06-01 03:22:24 +00:00
|
|
|
kfree_const(shrinker->name);
|
2022-07-20 14:47:55 +00:00
|
|
|
shrinker->name = NULL;
|
|
|
|
}
|
2022-06-01 03:22:24 +00:00
|
|
|
return err;
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
|
|
|
|
{
|
|
|
|
return __register_shrinker(shrinker);
|
|
|
|
}
|
|
|
|
#endif
|
2007-07-17 11:03:17 +00:00
|
|
|
EXPORT_SYMBOL(register_shrinker);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Remove one
|
|
|
|
*/
|
2007-07-17 11:03:17 +00:00
|
|
|
void unregister_shrinker(struct shrinker *shrinker)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2023-02-02 10:56:12 +00:00
|
|
|
struct dentry *debugfs_entry;
|
|
|
|
|
2021-05-05 01:36:29 +00:00
|
|
|
if (!(shrinker->flags & SHRINKER_REGISTERED))
|
2017-12-18 11:31:41 +00:00
|
|
|
return;
|
2021-05-05 01:36:29 +00:00
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
down_write(&shrinker_rwsem);
|
|
|
|
list_del(&shrinker->list);
|
2021-05-05 01:36:29 +00:00
|
|
|
shrinker->flags &= ~SHRINKER_REGISTERED;
|
|
|
|
if (shrinker->flags & SHRINKER_MEMCG_AWARE)
|
|
|
|
unregister_memcg_shrinker(shrinker);
|
2023-02-02 10:56:12 +00:00
|
|
|
debugfs_entry = shrinker_debugfs_remove(shrinker);
|
2005-04-16 22:20:36 +00:00
|
|
|
up_write(&shrinker_rwsem);
|
2021-05-05 01:36:29 +00:00
|
|
|
|
2023-02-02 10:56:12 +00:00
|
|
|
debugfs_remove_recursive(debugfs_entry);
|
|
|
|
|
2013-10-16 20:46:46 +00:00
|
|
|
kfree(shrinker->nr_deferred);
|
2017-12-18 11:31:41 +00:00
|
|
|
shrinker->nr_deferred = NULL;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2007-07-17 11:03:17 +00:00
|
|
|
EXPORT_SYMBOL(unregister_shrinker);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2021-04-15 11:56:23 +00:00
|
|
|
/**
|
|
|
|
* synchronize_shrinkers - Wait for all running shrinkers to complete.
|
|
|
|
*
|
|
|
|
* This is equivalent to calling unregister_shrink() and register_shrinker(),
|
|
|
|
* but atomically and with less overhead. This is useful to guarantee that all
|
|
|
|
* shrinker invocations have seen an update, before freeing memory, similar to
|
|
|
|
* rcu.
|
|
|
|
*/
|
|
|
|
void synchronize_shrinkers(void)
|
|
|
|
{
|
|
|
|
down_write(&shrinker_rwsem);
|
|
|
|
up_write(&shrinker_rwsem);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(synchronize_shrinkers);
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
#define SHRINK_BATCH 128
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
2015-02-12 22:58:54 +00:00
|
|
|
static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
|
mm: use sc->priority for slab shrink targets
Previously we were using the ratio of the number of lru pages scanned to
the number of eligible lru pages to determine the number of slab objects
to scan. The problem with this is that these two things have nothing to
do with each other, so in slab heavy work loads where there is little to
no page cache we can end up with the pages scanned being a very low
number. This means that we reclaim next to no slab pages and waste a
lot of time reclaiming small amounts of space.
Consider the following scenario, where we have the following values and
the rest of the memory usage is in slab
Active: 58840 kB
Inactive: 46860 kB
Every time we do a get_scan_count() we do this
scan = size >> sc->priority
where sc->priority starts at DEF_PRIORITY, which is 12. The first loop
through reclaim would result in a scan target of 2 pages to 11715 total
inactive pages, and 3 pages to 14710 total active pages. This is a
really really small target for a system that is entirely slab pages.
And this is super optimistic, this assumes we even get to scan these
pages. We don't increment sc->nr_scanned unless we 1) isolate the page,
which assumes it's not in use, and 2) can lock the page. Under pressure
these numbers could probably go down, I'm sure there's some random pages
from daemons that aren't actually in use, so the targets get even
smaller.
Instead use sc->priority in the same way we use it to determine scan
amounts for the lru's. This generally equates to pages. Consider the
following
slab_pages = (nr_objects * object_size) / PAGE_SIZE
What we would like to do is
scan = slab_pages >> sc->priority
but we don't know the number of slab pages each shrinker controls, only
the objects. However say that theoretically we knew how many pages a
shrinker controlled, we'd still have to convert this to objects, which
would look like the following
scan = shrinker_pages >> sc->priority
scan_objects = (PAGE_SIZE / object_size) * scan
or written another way
scan_objects = (shrinker_pages >> sc->priority) *
(PAGE_SIZE / object_size)
which can thus be written
scan_objects = ((shrinker_pages * PAGE_SIZE) / object_size) >>
sc->priority
which is just
scan_objects = nr_objects >> sc->priority
We don't need to know exactly how many pages each shrinker represents,
it's objects are all the information we need. Making this change allows
us to place an appropriate amount of pressure on the shrinker pools for
their relative size.
Link: http://lkml.kernel.org/r/1510780549-6812-1-git-send-email-josef@toxicpanda.com
Signed-off-by: Josef Bacik <jbacik@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Dave Chinner <david@fromorbit.com>
Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-02-01 00:16:26 +00:00
|
|
|
struct shrinker *shrinker, int priority)
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
{
|
|
|
|
unsigned long freed = 0;
|
|
|
|
unsigned long long delta;
|
|
|
|
long total_scan;
|
2014-04-03 21:47:32 +00:00
|
|
|
long freeable;
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
long nr;
|
|
|
|
long new_nr;
|
|
|
|
long batch_size = shrinker->batch ? shrinker->batch
|
|
|
|
: SHRINK_BATCH;
|
2016-12-13 00:41:50 +00:00
|
|
|
long scanned = 0, next_deferred;
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
2014-04-03 21:47:32 +00:00
|
|
|
freeable = shrinker->count_objects(shrinker, shrinkctl);
|
2018-08-17 22:48:21 +00:00
|
|
|
if (freeable == 0 || freeable == SHRINK_EMPTY)
|
|
|
|
return freeable;
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* copy the current shrinker scan count into a local variable
|
|
|
|
* and zero it so that other concurrent shrinker invocations
|
|
|
|
* don't also do this scanning work.
|
|
|
|
*/
|
2021-05-05 01:36:36 +00:00
|
|
|
nr = xchg_nr_deferred(shrinker, shrinkctl);
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
mm: zero-seek shrinkers
The page cache and most shrinkable slab caches hold data that has been
read from disk, but there are some caches that only cache CPU work, such
as the dentry and inode caches of procfs and sysfs, as well as the subset
of radix tree nodes that track non-resident page cache.
Currently, all these are shrunk at the same rate: using DEFAULT_SEEKS for
the shrinker's seeks setting tells the reclaim algorithm that for every
two page cache pages scanned it should scan one slab object.
This is a bogus setting. A virtual inode that required no IO to create is
not twice as valuable as a page cache page; shadow cache entries with
eviction distances beyond the size of memory aren't either.
In most cases, the behavior in practice is still fine. Such virtual
caches don't tend to grow and assert themselves aggressively, and usually
get picked up before they cause problems. But there are scenarios where
that's not true.
Our database workloads suffer from two of those. For one, their file
workingset is several times bigger than available memory, which has the
kernel aggressively create shadow page cache entries for the non-resident
parts of it. The workingset code does tell the VM that most of these are
expendable, but the VM ends up balancing them 2:1 to cache pages as per
the seeks setting. This is a huge waste of memory.
These workloads also deal with tens of thousands of open files and use
/proc for introspection, which ends up growing the proc_inode_cache to
absurdly large sizes - again at the cost of valuable cache space, which
isn't a reasonable trade-off, given that proc inodes can be re-created
without involving the disk.
This patch implements a "zero-seek" setting for shrinkers that results in
a target ratio of 0:1 between their objects and IO-backed caches. This
allows such virtual caches to grow when memory is available (they do
cache/avoid CPU work after all), but effectively disables them as soon as
IO-backed objects are under pressure.
It then switches the shrinkers for procfs and sysfs metadata, as well as
excess page cache shadow nodes, to the new zero-seek setting.
Link: http://lkml.kernel.org/r/20181009184732.762-5-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Domas Mituzas <dmituzas@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Rik van Riel <riel@surriel.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:06:42 +00:00
|
|
|
if (shrinker->seeks) {
|
|
|
|
delta = freeable >> priority;
|
|
|
|
delta *= 4;
|
|
|
|
do_div(delta, shrinker->seeks);
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* These objects don't require any IO to create. Trim
|
|
|
|
* them aggressively under memory pressure to keep
|
|
|
|
* them from causing refetches in the IO caches.
|
|
|
|
*/
|
|
|
|
delta = freeable / 2;
|
|
|
|
}
|
mm: slowly shrink slabs with a relatively small number of objects
9092c71bb724 ("mm: use sc->priority for slab shrink targets") changed the
way that the target slab pressure is calculated and made it
priority-based:
delta = freeable >> priority;
delta *= 4;
do_div(delta, shrinker->seeks);
The problem is that on a default priority (which is 12) no pressure is
applied at all, if the number of potentially reclaimable objects is less
than 4096 (1<<12).
This causes the last objects on slab caches of no longer used cgroups to
(almost) never get reclaimed. It's obviously a waste of memory.
It can be especially painful, if these stale objects are holding a
reference to a dying cgroup. Slab LRU lists are reparented on memcg
offlining, but corresponding objects are still holding a reference to the
dying cgroup. If we don't scan these objects, the dying cgroup can't go
away. Most likely, the parent cgroup hasn't any directly charged objects,
only remaining objects from dying children cgroups. So it can easily hold
a reference to hundreds of dying cgroups.
If there are no big spikes in memory pressure, and new memory cgroups are
created and destroyed periodically, this causes the number of dying
cgroups grow steadily, causing a slow-ish and hard-to-detect memory
"leak". It's not a real leak, as the memory can be eventually reclaimed,
but it could not happen in a real life at all. I've seen hosts with a
steadily climbing number of dying cgroups, which doesn't show any signs of
a decline in months, despite the host is loaded with a production
workload.
It is an obvious waste of memory, and to prevent it, let's apply a minimal
pressure even on small shrinker lists. E.g. if there are freeable
objects, let's scan at least min(freeable, scan_batch) objects.
This fix significantly improves a chance of a dying cgroup to be
reclaimed, and together with some previous patches stops the steady growth
of the dying cgroups number on some of our hosts.
Link: http://lkml.kernel.org/r/20180905230759.12236-1-guro@fb.com
Fixes: 9092c71bb724 ("mm: use sc->priority for slab shrink targets")
Signed-off-by: Roman Gushchin <guro@fb.com>
Acked-by: Rik van Riel <riel@surriel.com>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2018-09-20 19:22:46 +00:00
|
|
|
|
2021-05-05 01:36:45 +00:00
|
|
|
total_scan = nr >> priority;
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
total_scan += delta;
|
2021-05-05 01:36:45 +00:00
|
|
|
total_scan = min(total_scan, (2 * freeable));
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
|
|
|
trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
|
mm: use sc->priority for slab shrink targets
Previously we were using the ratio of the number of lru pages scanned to
the number of eligible lru pages to determine the number of slab objects
to scan. The problem with this is that these two things have nothing to
do with each other, so in slab heavy work loads where there is little to
no page cache we can end up with the pages scanned being a very low
number. This means that we reclaim next to no slab pages and waste a
lot of time reclaiming small amounts of space.
Consider the following scenario, where we have the following values and
the rest of the memory usage is in slab
Active: 58840 kB
Inactive: 46860 kB
Every time we do a get_scan_count() we do this
scan = size >> sc->priority
where sc->priority starts at DEF_PRIORITY, which is 12. The first loop
through reclaim would result in a scan target of 2 pages to 11715 total
inactive pages, and 3 pages to 14710 total active pages. This is a
really really small target for a system that is entirely slab pages.
And this is super optimistic, this assumes we even get to scan these
pages. We don't increment sc->nr_scanned unless we 1) isolate the page,
which assumes it's not in use, and 2) can lock the page. Under pressure
these numbers could probably go down, I'm sure there's some random pages
from daemons that aren't actually in use, so the targets get even
smaller.
Instead use sc->priority in the same way we use it to determine scan
amounts for the lru's. This generally equates to pages. Consider the
following
slab_pages = (nr_objects * object_size) / PAGE_SIZE
What we would like to do is
scan = slab_pages >> sc->priority
but we don't know the number of slab pages each shrinker controls, only
the objects. However say that theoretically we knew how many pages a
shrinker controlled, we'd still have to convert this to objects, which
would look like the following
scan = shrinker_pages >> sc->priority
scan_objects = (PAGE_SIZE / object_size) * scan
or written another way
scan_objects = (shrinker_pages >> sc->priority) *
(PAGE_SIZE / object_size)
which can thus be written
scan_objects = ((shrinker_pages * PAGE_SIZE) / object_size) >>
sc->priority
which is just
scan_objects = nr_objects >> sc->priority
We don't need to know exactly how many pages each shrinker represents,
it's objects are all the information we need. Making this change allows
us to place an appropriate amount of pressure on the shrinker pools for
their relative size.
Link: http://lkml.kernel.org/r/1510780549-6812-1-git-send-email-josef@toxicpanda.com
Signed-off-by: Josef Bacik <jbacik@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Dave Chinner <david@fromorbit.com>
Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-02-01 00:16:26 +00:00
|
|
|
freeable, delta, total_scan, priority);
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
mm: vmscan: shrink all slab objects if tight on memory
When reclaiming kmem, we currently don't scan slabs that have less than
batch_size objects (see shrink_slab_node()):
while (total_scan >= batch_size) {
shrinkctl->nr_to_scan = batch_size;
shrinker->scan_objects(shrinker, shrinkctl);
total_scan -= batch_size;
}
If there are only a few shrinkers available, such a behavior won't cause
any problems, because the batch_size is usually small, but if we have a
lot of slab shrinkers, which is perfectly possible since FS shrinkers
are now per-superblock, we can end up with hundreds of megabytes of
practically unreclaimable kmem objects. For instance, mounting a
thousand of ext2 FS images with a hundred of files in each and iterating
over all the files using du(1) will result in about 200 Mb of FS caches
that cannot be dropped even with the aid of the vm.drop_caches sysctl!
This problem was initially pointed out by Glauber Costa [*]. Glauber
proposed to fix it by making the shrink_slab() always take at least one
pass, to put it simply, turning the scan loop above to a do{}while()
loop. However, this proposal was rejected, because it could result in
more aggressive and frequent slab shrinking even under low memory
pressure when total_scan is naturally very small.
This patch is a slightly modified version of Glauber's approach.
Similarly to Glauber's patch, it makes shrink_slab() scan less than
batch_size objects, but only if the total number of objects we want to
scan (total_scan) is greater than the total number of objects available
(max_pass). Since total_scan is biased as half max_pass if the current
delta change is small:
if (delta < max_pass / 4)
total_scan = min(total_scan, max_pass / 2);
this is only possible if we are scanning at high prio. That said, this
patch shouldn't change the vmscan behaviour if the memory pressure is
low, but if we are tight on memory, we will do our best by trying to
reclaim all available objects, which sounds reasonable.
[*] http://www.spinics.net/lists/cgroups/msg06913.html
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Glauber Costa <glommer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-01-23 23:53:22 +00:00
|
|
|
/*
|
|
|
|
* Normally, we should not scan less than batch_size objects in one
|
|
|
|
* pass to avoid too frequent shrinker calls, but if the slab has less
|
|
|
|
* than batch_size objects in total and we are really tight on memory,
|
|
|
|
* we will try to reclaim all available objects, otherwise we can end
|
|
|
|
* up failing allocations although there are plenty of reclaimable
|
|
|
|
* objects spread over several slabs with usage less than the
|
|
|
|
* batch_size.
|
|
|
|
*
|
|
|
|
* We detect the "tight on memory" situations by looking at the total
|
|
|
|
* number of objects we want to scan (total_scan). If it is greater
|
2014-04-03 21:47:32 +00:00
|
|
|
* than the total number of objects on slab (freeable), we must be
|
mm: vmscan: shrink all slab objects if tight on memory
When reclaiming kmem, we currently don't scan slabs that have less than
batch_size objects (see shrink_slab_node()):
while (total_scan >= batch_size) {
shrinkctl->nr_to_scan = batch_size;
shrinker->scan_objects(shrinker, shrinkctl);
total_scan -= batch_size;
}
If there are only a few shrinkers available, such a behavior won't cause
any problems, because the batch_size is usually small, but if we have a
lot of slab shrinkers, which is perfectly possible since FS shrinkers
are now per-superblock, we can end up with hundreds of megabytes of
practically unreclaimable kmem objects. For instance, mounting a
thousand of ext2 FS images with a hundred of files in each and iterating
over all the files using du(1) will result in about 200 Mb of FS caches
that cannot be dropped even with the aid of the vm.drop_caches sysctl!
This problem was initially pointed out by Glauber Costa [*]. Glauber
proposed to fix it by making the shrink_slab() always take at least one
pass, to put it simply, turning the scan loop above to a do{}while()
loop. However, this proposal was rejected, because it could result in
more aggressive and frequent slab shrinking even under low memory
pressure when total_scan is naturally very small.
This patch is a slightly modified version of Glauber's approach.
Similarly to Glauber's patch, it makes shrink_slab() scan less than
batch_size objects, but only if the total number of objects we want to
scan (total_scan) is greater than the total number of objects available
(max_pass). Since total_scan is biased as half max_pass if the current
delta change is small:
if (delta < max_pass / 4)
total_scan = min(total_scan, max_pass / 2);
this is only possible if we are scanning at high prio. That said, this
patch shouldn't change the vmscan behaviour if the memory pressure is
low, but if we are tight on memory, we will do our best by trying to
reclaim all available objects, which sounds reasonable.
[*] http://www.spinics.net/lists/cgroups/msg06913.html
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Glauber Costa <glommer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-01-23 23:53:22 +00:00
|
|
|
* scanning at high prio and therefore should try to reclaim as much as
|
|
|
|
* possible.
|
|
|
|
*/
|
|
|
|
while (total_scan >= batch_size ||
|
2014-04-03 21:47:32 +00:00
|
|
|
total_scan >= freeable) {
|
2013-08-28 00:18:16 +00:00
|
|
|
unsigned long ret;
|
mm: vmscan: shrink all slab objects if tight on memory
When reclaiming kmem, we currently don't scan slabs that have less than
batch_size objects (see shrink_slab_node()):
while (total_scan >= batch_size) {
shrinkctl->nr_to_scan = batch_size;
shrinker->scan_objects(shrinker, shrinkctl);
total_scan -= batch_size;
}
If there are only a few shrinkers available, such a behavior won't cause
any problems, because the batch_size is usually small, but if we have a
lot of slab shrinkers, which is perfectly possible since FS shrinkers
are now per-superblock, we can end up with hundreds of megabytes of
practically unreclaimable kmem objects. For instance, mounting a
thousand of ext2 FS images with a hundred of files in each and iterating
over all the files using du(1) will result in about 200 Mb of FS caches
that cannot be dropped even with the aid of the vm.drop_caches sysctl!
This problem was initially pointed out by Glauber Costa [*]. Glauber
proposed to fix it by making the shrink_slab() always take at least one
pass, to put it simply, turning the scan loop above to a do{}while()
loop. However, this proposal was rejected, because it could result in
more aggressive and frequent slab shrinking even under low memory
pressure when total_scan is naturally very small.
This patch is a slightly modified version of Glauber's approach.
Similarly to Glauber's patch, it makes shrink_slab() scan less than
batch_size objects, but only if the total number of objects we want to
scan (total_scan) is greater than the total number of objects available
(max_pass). Since total_scan is biased as half max_pass if the current
delta change is small:
if (delta < max_pass / 4)
total_scan = min(total_scan, max_pass / 2);
this is only possible if we are scanning at high prio. That said, this
patch shouldn't change the vmscan behaviour if the memory pressure is
low, but if we are tight on memory, we will do our best by trying to
reclaim all available objects, which sounds reasonable.
[*] http://www.spinics.net/lists/cgroups/msg06913.html
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Glauber Costa <glommer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-01-23 23:53:22 +00:00
|
|
|
unsigned long nr_to_scan = min(batch_size, total_scan);
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
mm: vmscan: shrink all slab objects if tight on memory
When reclaiming kmem, we currently don't scan slabs that have less than
batch_size objects (see shrink_slab_node()):
while (total_scan >= batch_size) {
shrinkctl->nr_to_scan = batch_size;
shrinker->scan_objects(shrinker, shrinkctl);
total_scan -= batch_size;
}
If there are only a few shrinkers available, such a behavior won't cause
any problems, because the batch_size is usually small, but if we have a
lot of slab shrinkers, which is perfectly possible since FS shrinkers
are now per-superblock, we can end up with hundreds of megabytes of
practically unreclaimable kmem objects. For instance, mounting a
thousand of ext2 FS images with a hundred of files in each and iterating
over all the files using du(1) will result in about 200 Mb of FS caches
that cannot be dropped even with the aid of the vm.drop_caches sysctl!
This problem was initially pointed out by Glauber Costa [*]. Glauber
proposed to fix it by making the shrink_slab() always take at least one
pass, to put it simply, turning the scan loop above to a do{}while()
loop. However, this proposal was rejected, because it could result in
more aggressive and frequent slab shrinking even under low memory
pressure when total_scan is naturally very small.
This patch is a slightly modified version of Glauber's approach.
Similarly to Glauber's patch, it makes shrink_slab() scan less than
batch_size objects, but only if the total number of objects we want to
scan (total_scan) is greater than the total number of objects available
(max_pass). Since total_scan is biased as half max_pass if the current
delta change is small:
if (delta < max_pass / 4)
total_scan = min(total_scan, max_pass / 2);
this is only possible if we are scanning at high prio. That said, this
patch shouldn't change the vmscan behaviour if the memory pressure is
low, but if we are tight on memory, we will do our best by trying to
reclaim all available objects, which sounds reasonable.
[*] http://www.spinics.net/lists/cgroups/msg06913.html
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Glauber Costa <glommer@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-01-23 23:53:22 +00:00
|
|
|
shrinkctl->nr_to_scan = nr_to_scan;
|
2017-09-06 23:19:26 +00:00
|
|
|
shrinkctl->nr_scanned = nr_to_scan;
|
2013-08-28 00:18:16 +00:00
|
|
|
ret = shrinker->scan_objects(shrinker, shrinkctl);
|
|
|
|
if (ret == SHRINK_STOP)
|
|
|
|
break;
|
|
|
|
freed += ret;
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
2017-09-06 23:19:26 +00:00
|
|
|
count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
|
|
|
|
total_scan -= shrinkctl->nr_scanned;
|
|
|
|
scanned += shrinkctl->nr_scanned;
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
|
|
|
cond_resched();
|
|
|
|
}
|
|
|
|
|
2021-05-05 01:36:45 +00:00
|
|
|
/*
|
|
|
|
* The deferred work is increased by any new work (delta) that wasn't
|
|
|
|
* done, decreased by old deferred work that was done now.
|
|
|
|
*
|
|
|
|
* And it is capped to two times of the freeable items.
|
|
|
|
*/
|
|
|
|
next_deferred = max_t(long, (nr + delta - scanned), 0);
|
|
|
|
next_deferred = min(next_deferred, (2 * freeable));
|
|
|
|
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
/*
|
|
|
|
* move the unused scan count back into the shrinker in a
|
2021-05-05 01:36:36 +00:00
|
|
|
* manner that handles concurrent updates.
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
*/
|
2021-05-05 01:36:36 +00:00
|
|
|
new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
|
mm: vmscan: use nid from shrink_control for tracepoint
Patch series "Make shrinker's nr_deferred memcg aware", v10.
Recently huge amount one-off slab drop was seen on some vfs metadata
heavy workloads, it turned out there were huge amount accumulated
nr_deferred objects seen by the shrinker.
On our production machine, I saw absurd number of nr_deferred shown as
the below tracing result:
<...>-48776 [032] .... 27970562.458916: mm_shrink_slab_start:
super_cache_scan+0x0/0x1a0 ffff9a83046f3458: nid: 0 objects to shrink
2531805877005 gfp_flags GFP_HIGHUSER_MOVABLE pgs_scanned 32 lru_pgs
9300 cache items 1667 delta 11 total_scan 833
There are 2.5 trillion deferred objects on one node, assuming all of them
are dentry (192 bytes per object), so the total size of deferred on one
node is ~480TB. It is definitely ridiculous.
I managed to reproduce this problem with kernel build workload plus
negative dentry generator.
First step, run the below kernel build test script:
NR_CPUS=`cat /proc/cpuinfo | grep -e processor | wc -l`
cd /root/Buildarea/linux-stable
for i in `seq 1500`; do
cgcreate -g memory:kern_build
echo 4G > /sys/fs/cgroup/memory/kern_build/memory.limit_in_bytes
echo 3 > /proc/sys/vm/drop_caches
cgexec -g memory:kern_build make clean > /dev/null 2>&1
cgexec -g memory:kern_build make -j$NR_CPUS > /dev/null 2>&1
cgdelete -g memory:kern_build
done
Then run the below negative dentry generator script:
NR_CPUS=`cat /proc/cpuinfo | grep -e processor | wc -l`
mkdir /sys/fs/cgroup/memory/test
echo $$ > /sys/fs/cgroup/memory/test/tasks
for i in `seq $NR_CPUS`; do
while true; do
FILE=`head /dev/urandom | tr -dc A-Za-z0-9 | head -c 64`
cat $FILE 2>/dev/null
done &
done
Then kswapd will shrink half of dentry cache in just one loop as the below
tracing result showed:
kswapd0-475 [028] .... 305968.252561: mm_shrink_slab_start: super_cache_scan+0x0/0x190 0000000024acf00c: nid: 0 objects to shrink 4994376020 gfp_flags GFP_KERNEL cache items 93689873 delta 45746 total_scan 46844936 priority 12
kswapd0-475 [021] .... 306013.099399: mm_shrink_slab_end: super_cache_scan+0x0/0x190 0000000024acf00c: nid: 0 unused scan count 4994376020 new scan count 4947576838 total_scan 8 last shrinker return val 46844928
There were huge number of deferred objects before the shrinker was called,
the behavior does match the code but it might be not desirable from the
user's stand of point.
The excessive amount of nr_deferred might be accumulated due to various
reasons, for example:
* GFP_NOFS allocation
* Significant times of small amount scan (< scan_batch, 1024 for vfs
metadata)
However the LRUs of slabs are per memcg (memcg-aware shrinkers) but the
deferred objects is per shrinker, this may have some bad effects:
* Poor isolation among memcgs. Some memcgs which happen to have
frequent limit reclaim may get nr_deferred accumulated to a huge number,
then other innocent memcgs may take the fall. In our case the main
workload was hit.
* Unbounded deferred objects. There is no cap for deferred objects, it
can outgrow ridiculously as the tracing result showed.
* Easy to get out of control. Although shrinkers take into account
deferred objects, but it can go out of control easily. One
misconfigured memcg could incur absurd amount of deferred objects in a
period of time.
* Sort of reclaim problems, i.e. over reclaim, long reclaim latency,
etc. There may be hundred GB slab caches for vfe metadata heavy
workload, shrink half of them may take minutes. We observed latency
spike due to the prolonged reclaim.
These issues also have been discussed in
https://lore.kernel.org/linux-mm/20200916185823.5347-1-shy828301@gmail.com/.
The patchset is the outcome of that discussion.
So this patchset makes nr_deferred per-memcg to tackle the problem. It
does:
* Have memcg_shrinker_deferred per memcg per node, just like what
shrinker_map does. Instead it is an atomic_long_t array, each element
represent one shrinker even though the shrinker is not memcg aware, this
simplifies the implementation. For memcg aware shrinkers, the deferred
objects are just accumulated to its own memcg. The shrinkers just see
nr_deferred from its own memcg. Non memcg aware shrinkers still use
global nr_deferred from struct shrinker.
* Once the memcg is offlined, its nr_deferred will be reparented to its
parent along with LRUs.
* The root memcg has memcg_shrinker_deferred array too. It simplifies
the handling of reparenting to root memcg.
* Cap nr_deferred to 2x of the length of lru. The idea is borrowed from
Dave Chinner's series
(https://lore.kernel.org/linux-xfs/20191031234618.15403-1-david@fromorbit.com/)
The downside is each memcg has to allocate extra memory to store the
nr_deferred array. On our production environment, there are typically
around 40 shrinkers, so each memcg needs ~320 bytes. 10K memcgs would
need ~3.2MB memory. It seems fine.
We have been running the patched kernel on some hosts of our fleet (test
and production) for months, it works very well. The monitor data shows
the working set is sustained as expected.
This patch (of 13):
The tracepoint's nid should show what node the shrink happens on, the
start tracepoint uses nid from shrinkctl, but the nid might be set to 0
before end tracepoint if the shrinker is not NUMA aware, so the tracing
log may show the shrink happens on one node but end up on the other node.
It seems confusing. And the following patch will remove using nid
directly in do_shrink_slab(), this patch also helps cleanup the code.
Link: https://lkml.kernel.org/r/20210311190845.9708-1-shy828301@gmail.com
Link: https://lkml.kernel.org/r/20210311190845.9708-2-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Roman Gushchin <guro@fb.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:36:08 +00:00
|
|
|
trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
|
vmscan: per-node deferred work
The list_lru infrastructure already keeps per-node LRU lists in its
node-specific list_lru_node arrays and provide us with a per-node API, and
the shrinkers are properly equiped with node information. This means that
we can now focus our shrinking effort in a single node, but the work that
is deferred from one run to another is kept global at nr_in_batch. Work
can be deferred, for instance, during direct reclaim under a GFP_NOFS
allocation, where situation, all the filesystem shrinkers will be
prevented from running and accumulate in nr_in_batch the amount of work
they should have done, but could not.
This creates an impedance problem, where upon node pressure, work deferred
will accumulate and end up being flushed in other nodes. The problem we
describe is particularly harmful in big machines, where many nodes can
accumulate at the same time, all adding to the global counter nr_in_batch.
As we accumulate more and more, we start to ask for the caches to flush
even bigger numbers. The result is that the caches are depleted and do
not stabilize. To achieve stable steady state behavior, we need to tackle
it differently.
In this patch we keep the deferred count per-node, in the new array
nr_deferred[] (the name is also a bit more descriptive) and will never
accumulate that to other nodes.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 00:18:04 +00:00
|
|
|
return freed;
|
2011-05-25 00:12:27 +00:00
|
|
|
}
|
|
|
|
|
2019-09-23 22:38:12 +00:00
|
|
|
#ifdef CONFIG_MEMCG
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
|
|
|
|
struct mem_cgroup *memcg, int priority)
|
|
|
|
{
|
2021-05-05 01:36:23 +00:00
|
|
|
struct shrinker_info *info;
|
2018-10-05 22:52:10 +00:00
|
|
|
unsigned long ret, freed = 0;
|
|
|
|
int i;
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
|
2019-09-23 22:38:12 +00:00
|
|
|
if (!mem_cgroup_online(memcg))
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (!down_read_trylock(&shrinker_rwsem))
|
|
|
|
return 0;
|
|
|
|
|
2021-05-05 01:36:26 +00:00
|
|
|
info = shrinker_info_protected(memcg, nid);
|
2021-05-05 01:36:23 +00:00
|
|
|
if (unlikely(!info))
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
goto unlock;
|
|
|
|
|
2021-05-05 01:36:23 +00:00
|
|
|
for_each_set_bit(i, info->map, shrinker_nr_max) {
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
struct shrink_control sc = {
|
|
|
|
.gfp_mask = gfp_mask,
|
|
|
|
.nid = nid,
|
|
|
|
.memcg = memcg,
|
|
|
|
};
|
|
|
|
struct shrinker *shrinker;
|
|
|
|
|
|
|
|
shrinker = idr_find(&shrinker_idr, i);
|
2021-05-05 01:36:29 +00:00
|
|
|
if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
|
mm: use special value SHRINKER_REGISTERING instead of list_empty() check
The patch introduces a special value SHRINKER_REGISTERING to use instead
of list_empty() to differ a registering shrinker from unregistered
shrinker. Why we need that at all?
Shrinker registration is split in two parts. The first one is
prealloc_shrinker(), which allocates shrinker memory and reserves ID in
shrinker_idr. This function can fail. The second is
register_shrinker_prepared(), and it finalizes the registration. This
function actually makes shrinker available to be used from
shrink_slab(), and it can't fail.
One shrinker may be based on more then one LRU lists. So, we never
clear the bit in memcg shrinker maps, when (one of) corresponding LRU
list becomes empty, since other LRU lists may be not empty. See
superblock shrinker for example: it is based on two LRU lists:
s_inode_lru and s_dentry_lru. We do not want to clear shrinker bit,
when there are no inodes in s_inode_lru, as s_dentry_lru may contain
dentries.
Instead of that, we use special algorithm to detect shrinkers having no
elements at all its LRU lists, and this is made in shrink_slab_memcg().
See the comment in this function for the details.
Also, in shrink_slab_memcg() we clear shrinker bit in the map, when we
meet unregistered shrinker (bit is set, while there is no a shrinker in
IDR). Otherwise, we would have done that at the moment of shrinker
unregistration for all memcgs (and this looks worse, since iteration
over all memcg may take much time). Also this would have imposed
restrictions on shrinker unregistration order for its users: they would
have had to guarantee, there are no new elements after
unregister_shrinker() (otherwise, a new added element would have set a
bit).
So, if we meet a set bit in map and no shrinker in IDR when we're
iterating over the map in shrink_slab_memcg(), this means the
corresponding shrinker is unregistered, and we must clear the bit.
Another case is shrinker registration. We want two things there:
1) do_shrink_slab() can be called only for completely registered
shrinkers;
2) shrinker internal lists may be populated in any order with
register_shrinker_prepared() (let's talk on the example with sb). Both
of:
a)list_lru_add(&inode->i_sb->s_inode_lru, &inode->i_lru); [cpu0]
memcg_set_shrinker_bit(); [cpu0]
...
register_shrinker_prepared(); [cpu1]
and
b)register_shrinker_prepared(); [cpu0]
...
list_lru_add(&inode->i_sb->s_inode_lru, &inode->i_lru); [cpu1]
memcg_set_shrinker_bit(); [cpu1]
are legitimate. We don't want to impose restriction here and to
force people to use only (b) variant. We don't want to force people to
care, there is no elements in LRU lists before the shrinker is
completely registered. Internal users of LRU lists and shrinker code
are two different subsystems, and they have to be closed in themselves
each other.
In (a) case we have the bit set before shrinker is completely
registered. We don't want do_shrink_slab() is called at this moment, so
we have to detect such the registering shrinkers.
Before this patch list_empty() (shrinker is not linked to the list)
check was used for that. So, in (a) there could be a bit set, but we
don't call do_shrink_slab() unless shrinker is linked to the list. It's
just an indicator, I just overloaded linking to the list.
This was not the best solution, since it's better not to touch the
shrinker memory from shrink_slab_memcg() before it's completely
registered (this also will be useful in the future to make shrink_slab()
completely lockless).
So, this patch introduces better way to detect registering shrinker,
which allows not to dereference shrinker memory. It's just a ~0UL
value, which we insert into the IDR during ID allocation. After
shrinker is ready to be used, we insert actual shrinker pointer in the
IDR, and it becomes available to shrink_slab_memcg().
We can't use NULL instead of this new value for this purpose as:
shrink_slab_memcg() already uses NULL to detect unregistered shrinkers,
and we don't want the function sees NULL and clears the bit, otherwise
(a) won't work.
This is the only thing the patch makes: the better way to detect
registering shrinker. Nothing else this patch makes.
Also this gives a better assembler, but it's minor side of the patch:
Before:
callq <idr_find>
mov %rax,%r15
test %rax,%rax
je <shrink_slab_memcg+0x1d5>
mov 0x20(%rax),%rax
lea 0x20(%r15),%rdx
cmp %rax,%rdx
je <shrink_slab_memcg+0xbd>
mov 0x8(%rsp),%edx
mov %r15,%rsi
lea 0x10(%rsp),%rdi
callq <do_shrink_slab>
After:
callq <idr_find>
mov %rax,%r15
lea -0x1(%rax),%rax
cmp $0xfffffffffffffffd,%rax
ja <shrink_slab_memcg+0x1cd>
mov 0x8(%rsp),%edx
mov %r15,%rsi
lea 0x10(%rsp),%rdi
callq ffffffff810cefd0 <do_shrink_slab>
[ktkhai@virtuozzo.com: add #ifdef CONFIG_MEMCG_KMEM around idr_replace()]
Link: http://lkml.kernel.org/r/758b8fec-7573-47eb-b26a-7b2847ae7b8c@virtuozzo.com
Link: http://lkml.kernel.org/r/153355467546.11522.4518015068123480218.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Josef Bacik <jbacik@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:34 +00:00
|
|
|
if (!shrinker)
|
2021-05-05 01:36:23 +00:00
|
|
|
clear_bit(i, info->map);
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2019-09-23 22:38:12 +00:00
|
|
|
/* Call non-slab shrinkers even though kmem is disabled */
|
2023-02-13 19:29:22 +00:00
|
|
|
if (!memcg_kmem_online() &&
|
2019-09-23 22:38:12 +00:00
|
|
|
!(shrinker->flags & SHRINKER_NONSLAB))
|
|
|
|
continue;
|
|
|
|
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
ret = do_shrink_slab(&sc, shrinker, priority);
|
mm/vmscan.c: clear shrinker bit if there are no objects related to memcg
To avoid further unneed calls of do_shrink_slab() for shrinkers, which
already do not have any charged objects in a memcg, their bits have to
be cleared.
This patch introduces a lockless mechanism to do that without races
without parallel list lru add. After do_shrink_slab() returns
SHRINK_EMPTY the first time, we clear the bit and call it once again.
Then we restore the bit, if the new return value is different.
Note, that single smp_mb__after_atomic() in shrink_slab_memcg() covers
two situations:
1)list_lru_add() shrink_slab_memcg
list_add_tail() for_each_set_bit() <--- read bit
do_shrink_slab() <--- missed list update (no barrier)
<MB> <MB>
set_bit() do_shrink_slab() <--- seen list update
This situation, when the first do_shrink_slab() sees set bit, but it
doesn't see list update (i.e., race with the first element queueing), is
rare. So we don't add <MB> before the first call of do_shrink_slab()
instead of this to do not slow down generic case. Also, it's need the
second call as seen in below in (2).
2)list_lru_add() shrink_slab_memcg()
list_add_tail() ...
set_bit() ...
... for_each_set_bit()
do_shrink_slab() do_shrink_slab()
clear_bit() ...
... ...
list_lru_add() ...
list_add_tail() clear_bit()
<MB> <MB>
set_bit() do_shrink_slab()
The barriers guarantee that the second do_shrink_slab() in the right
side task sees list update if really cleared the bit. This case is
drawn in the code comment.
[Results/performance of the patchset]
After the whole patchset applied the below test shows signify increase
of performance:
$echo 1 > /sys/fs/cgroup/memory/memory.use_hierarchy
$mkdir /sys/fs/cgroup/memory/ct
$echo 4000M > /sys/fs/cgroup/memory/ct/memory.kmem.limit_in_bytes
$for i in `seq 0 4000`; do mkdir /sys/fs/cgroup/memory/ct/$i;
echo $$ > /sys/fs/cgroup/memory/ct/$i/cgroup.procs;
mkdir -p s/$i; mount -t tmpfs $i s/$i;
touch s/$i/file; done
Then, 5 sequential calls of drop caches:
$time echo 3 > /proc/sys/vm/drop_caches
1)Before:
0.00user 13.78system 0:13.78elapsed 99%CPU
0.00user 5.59system 0:05.60elapsed 99%CPU
0.00user 5.48system 0:05.48elapsed 99%CPU
0.00user 8.35system 0:08.35elapsed 99%CPU
0.00user 8.34system 0:08.35elapsed 99%CPU
2)After
0.00user 1.10system 0:01.10elapsed 99%CPU
0.00user 0.00system 0:00.01elapsed 64%CPU
0.00user 0.01system 0:00.01elapsed 82%CPU
0.00user 0.00system 0:00.01elapsed 64%CPU
0.00user 0.01system 0:00.01elapsed 82%CPU
The results show the performance increases at least in 548 times.
Shakeel Butt tested this patchset with fork-bomb on his configuration:
> I created 255 memcgs, 255 ext4 mounts and made each memcg create a
> file containing few KiBs on corresponding mount. Then in a separate
> memcg of 200 MiB limit ran a fork-bomb.
>
> I ran the "perf record -ag -- sleep 60" and below are the results:
>
> Without the patch series:
> Samples: 4M of event 'cycles', Event count (approx.): 3279403076005
> + 36.40% fb.sh [kernel.kallsyms] [k] shrink_slab
> + 18.97% fb.sh [kernel.kallsyms] [k] list_lru_count_one
> + 6.75% fb.sh [kernel.kallsyms] [k] super_cache_count
> + 0.49% fb.sh [kernel.kallsyms] [k] down_read_trylock
> + 0.44% fb.sh [kernel.kallsyms] [k] mem_cgroup_iter
> + 0.27% fb.sh [kernel.kallsyms] [k] up_read
> + 0.21% fb.sh [kernel.kallsyms] [k] osq_lock
> + 0.13% fb.sh [kernel.kallsyms] [k] shmem_unused_huge_count
> + 0.08% fb.sh [kernel.kallsyms] [k] shrink_node_memcg
> + 0.08% fb.sh [kernel.kallsyms] [k] shrink_node
>
> With the patch series:
> Samples: 4M of event 'cycles', Event count (approx.): 2756866824946
> + 47.49% fb.sh [kernel.kallsyms] [k] down_read_trylock
> + 30.72% fb.sh [kernel.kallsyms] [k] up_read
> + 9.51% fb.sh [kernel.kallsyms] [k] mem_cgroup_iter
> + 1.69% fb.sh [kernel.kallsyms] [k] shrink_node_memcg
> + 1.35% fb.sh [kernel.kallsyms] [k] mem_cgroup_protected
> + 1.05% fb.sh [kernel.kallsyms] [k] queued_spin_lock_slowpath
> + 0.85% fb.sh [kernel.kallsyms] [k] _raw_spin_lock
> + 0.78% fb.sh [kernel.kallsyms] [k] lruvec_lru_size
> + 0.57% fb.sh [kernel.kallsyms] [k] shrink_node
> + 0.54% fb.sh [kernel.kallsyms] [k] queue_work_on
> + 0.46% fb.sh [kernel.kallsyms] [k] shrink_slab_memcg
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112561772.4097.11011071937553113003.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063070859.1818.11870882950920963480.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:25 +00:00
|
|
|
if (ret == SHRINK_EMPTY) {
|
2021-05-05 01:36:23 +00:00
|
|
|
clear_bit(i, info->map);
|
mm/vmscan.c: clear shrinker bit if there are no objects related to memcg
To avoid further unneed calls of do_shrink_slab() for shrinkers, which
already do not have any charged objects in a memcg, their bits have to
be cleared.
This patch introduces a lockless mechanism to do that without races
without parallel list lru add. After do_shrink_slab() returns
SHRINK_EMPTY the first time, we clear the bit and call it once again.
Then we restore the bit, if the new return value is different.
Note, that single smp_mb__after_atomic() in shrink_slab_memcg() covers
two situations:
1)list_lru_add() shrink_slab_memcg
list_add_tail() for_each_set_bit() <--- read bit
do_shrink_slab() <--- missed list update (no barrier)
<MB> <MB>
set_bit() do_shrink_slab() <--- seen list update
This situation, when the first do_shrink_slab() sees set bit, but it
doesn't see list update (i.e., race with the first element queueing), is
rare. So we don't add <MB> before the first call of do_shrink_slab()
instead of this to do not slow down generic case. Also, it's need the
second call as seen in below in (2).
2)list_lru_add() shrink_slab_memcg()
list_add_tail() ...
set_bit() ...
... for_each_set_bit()
do_shrink_slab() do_shrink_slab()
clear_bit() ...
... ...
list_lru_add() ...
list_add_tail() clear_bit()
<MB> <MB>
set_bit() do_shrink_slab()
The barriers guarantee that the second do_shrink_slab() in the right
side task sees list update if really cleared the bit. This case is
drawn in the code comment.
[Results/performance of the patchset]
After the whole patchset applied the below test shows signify increase
of performance:
$echo 1 > /sys/fs/cgroup/memory/memory.use_hierarchy
$mkdir /sys/fs/cgroup/memory/ct
$echo 4000M > /sys/fs/cgroup/memory/ct/memory.kmem.limit_in_bytes
$for i in `seq 0 4000`; do mkdir /sys/fs/cgroup/memory/ct/$i;
echo $$ > /sys/fs/cgroup/memory/ct/$i/cgroup.procs;
mkdir -p s/$i; mount -t tmpfs $i s/$i;
touch s/$i/file; done
Then, 5 sequential calls of drop caches:
$time echo 3 > /proc/sys/vm/drop_caches
1)Before:
0.00user 13.78system 0:13.78elapsed 99%CPU
0.00user 5.59system 0:05.60elapsed 99%CPU
0.00user 5.48system 0:05.48elapsed 99%CPU
0.00user 8.35system 0:08.35elapsed 99%CPU
0.00user 8.34system 0:08.35elapsed 99%CPU
2)After
0.00user 1.10system 0:01.10elapsed 99%CPU
0.00user 0.00system 0:00.01elapsed 64%CPU
0.00user 0.01system 0:00.01elapsed 82%CPU
0.00user 0.00system 0:00.01elapsed 64%CPU
0.00user 0.01system 0:00.01elapsed 82%CPU
The results show the performance increases at least in 548 times.
Shakeel Butt tested this patchset with fork-bomb on his configuration:
> I created 255 memcgs, 255 ext4 mounts and made each memcg create a
> file containing few KiBs on corresponding mount. Then in a separate
> memcg of 200 MiB limit ran a fork-bomb.
>
> I ran the "perf record -ag -- sleep 60" and below are the results:
>
> Without the patch series:
> Samples: 4M of event 'cycles', Event count (approx.): 3279403076005
> + 36.40% fb.sh [kernel.kallsyms] [k] shrink_slab
> + 18.97% fb.sh [kernel.kallsyms] [k] list_lru_count_one
> + 6.75% fb.sh [kernel.kallsyms] [k] super_cache_count
> + 0.49% fb.sh [kernel.kallsyms] [k] down_read_trylock
> + 0.44% fb.sh [kernel.kallsyms] [k] mem_cgroup_iter
> + 0.27% fb.sh [kernel.kallsyms] [k] up_read
> + 0.21% fb.sh [kernel.kallsyms] [k] osq_lock
> + 0.13% fb.sh [kernel.kallsyms] [k] shmem_unused_huge_count
> + 0.08% fb.sh [kernel.kallsyms] [k] shrink_node_memcg
> + 0.08% fb.sh [kernel.kallsyms] [k] shrink_node
>
> With the patch series:
> Samples: 4M of event 'cycles', Event count (approx.): 2756866824946
> + 47.49% fb.sh [kernel.kallsyms] [k] down_read_trylock
> + 30.72% fb.sh [kernel.kallsyms] [k] up_read
> + 9.51% fb.sh [kernel.kallsyms] [k] mem_cgroup_iter
> + 1.69% fb.sh [kernel.kallsyms] [k] shrink_node_memcg
> + 1.35% fb.sh [kernel.kallsyms] [k] mem_cgroup_protected
> + 1.05% fb.sh [kernel.kallsyms] [k] queued_spin_lock_slowpath
> + 0.85% fb.sh [kernel.kallsyms] [k] _raw_spin_lock
> + 0.78% fb.sh [kernel.kallsyms] [k] lruvec_lru_size
> + 0.57% fb.sh [kernel.kallsyms] [k] shrink_node
> + 0.54% fb.sh [kernel.kallsyms] [k] queue_work_on
> + 0.46% fb.sh [kernel.kallsyms] [k] shrink_slab_memcg
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112561772.4097.11011071937553113003.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063070859.1818.11870882950920963480.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:25 +00:00
|
|
|
/*
|
|
|
|
* After the shrinker reported that it had no objects to
|
|
|
|
* free, but before we cleared the corresponding bit in
|
|
|
|
* the memcg shrinker map, a new object might have been
|
|
|
|
* added. To make sure, we have the bit set in this
|
|
|
|
* case, we invoke the shrinker one more time and reset
|
|
|
|
* the bit if it reports that it is not empty anymore.
|
|
|
|
* The memory barrier here pairs with the barrier in
|
2021-05-05 01:36:11 +00:00
|
|
|
* set_shrinker_bit():
|
mm/vmscan.c: clear shrinker bit if there are no objects related to memcg
To avoid further unneed calls of do_shrink_slab() for shrinkers, which
already do not have any charged objects in a memcg, their bits have to
be cleared.
This patch introduces a lockless mechanism to do that without races
without parallel list lru add. After do_shrink_slab() returns
SHRINK_EMPTY the first time, we clear the bit and call it once again.
Then we restore the bit, if the new return value is different.
Note, that single smp_mb__after_atomic() in shrink_slab_memcg() covers
two situations:
1)list_lru_add() shrink_slab_memcg
list_add_tail() for_each_set_bit() <--- read bit
do_shrink_slab() <--- missed list update (no barrier)
<MB> <MB>
set_bit() do_shrink_slab() <--- seen list update
This situation, when the first do_shrink_slab() sees set bit, but it
doesn't see list update (i.e., race with the first element queueing), is
rare. So we don't add <MB> before the first call of do_shrink_slab()
instead of this to do not slow down generic case. Also, it's need the
second call as seen in below in (2).
2)list_lru_add() shrink_slab_memcg()
list_add_tail() ...
set_bit() ...
... for_each_set_bit()
do_shrink_slab() do_shrink_slab()
clear_bit() ...
... ...
list_lru_add() ...
list_add_tail() clear_bit()
<MB> <MB>
set_bit() do_shrink_slab()
The barriers guarantee that the second do_shrink_slab() in the right
side task sees list update if really cleared the bit. This case is
drawn in the code comment.
[Results/performance of the patchset]
After the whole patchset applied the below test shows signify increase
of performance:
$echo 1 > /sys/fs/cgroup/memory/memory.use_hierarchy
$mkdir /sys/fs/cgroup/memory/ct
$echo 4000M > /sys/fs/cgroup/memory/ct/memory.kmem.limit_in_bytes
$for i in `seq 0 4000`; do mkdir /sys/fs/cgroup/memory/ct/$i;
echo $$ > /sys/fs/cgroup/memory/ct/$i/cgroup.procs;
mkdir -p s/$i; mount -t tmpfs $i s/$i;
touch s/$i/file; done
Then, 5 sequential calls of drop caches:
$time echo 3 > /proc/sys/vm/drop_caches
1)Before:
0.00user 13.78system 0:13.78elapsed 99%CPU
0.00user 5.59system 0:05.60elapsed 99%CPU
0.00user 5.48system 0:05.48elapsed 99%CPU
0.00user 8.35system 0:08.35elapsed 99%CPU
0.00user 8.34system 0:08.35elapsed 99%CPU
2)After
0.00user 1.10system 0:01.10elapsed 99%CPU
0.00user 0.00system 0:00.01elapsed 64%CPU
0.00user 0.01system 0:00.01elapsed 82%CPU
0.00user 0.00system 0:00.01elapsed 64%CPU
0.00user 0.01system 0:00.01elapsed 82%CPU
The results show the performance increases at least in 548 times.
Shakeel Butt tested this patchset with fork-bomb on his configuration:
> I created 255 memcgs, 255 ext4 mounts and made each memcg create a
> file containing few KiBs on corresponding mount. Then in a separate
> memcg of 200 MiB limit ran a fork-bomb.
>
> I ran the "perf record -ag -- sleep 60" and below are the results:
>
> Without the patch series:
> Samples: 4M of event 'cycles', Event count (approx.): 3279403076005
> + 36.40% fb.sh [kernel.kallsyms] [k] shrink_slab
> + 18.97% fb.sh [kernel.kallsyms] [k] list_lru_count_one
> + 6.75% fb.sh [kernel.kallsyms] [k] super_cache_count
> + 0.49% fb.sh [kernel.kallsyms] [k] down_read_trylock
> + 0.44% fb.sh [kernel.kallsyms] [k] mem_cgroup_iter
> + 0.27% fb.sh [kernel.kallsyms] [k] up_read
> + 0.21% fb.sh [kernel.kallsyms] [k] osq_lock
> + 0.13% fb.sh [kernel.kallsyms] [k] shmem_unused_huge_count
> + 0.08% fb.sh [kernel.kallsyms] [k] shrink_node_memcg
> + 0.08% fb.sh [kernel.kallsyms] [k] shrink_node
>
> With the patch series:
> Samples: 4M of event 'cycles', Event count (approx.): 2756866824946
> + 47.49% fb.sh [kernel.kallsyms] [k] down_read_trylock
> + 30.72% fb.sh [kernel.kallsyms] [k] up_read
> + 9.51% fb.sh [kernel.kallsyms] [k] mem_cgroup_iter
> + 1.69% fb.sh [kernel.kallsyms] [k] shrink_node_memcg
> + 1.35% fb.sh [kernel.kallsyms] [k] mem_cgroup_protected
> + 1.05% fb.sh [kernel.kallsyms] [k] queued_spin_lock_slowpath
> + 0.85% fb.sh [kernel.kallsyms] [k] _raw_spin_lock
> + 0.78% fb.sh [kernel.kallsyms] [k] lruvec_lru_size
> + 0.57% fb.sh [kernel.kallsyms] [k] shrink_node
> + 0.54% fb.sh [kernel.kallsyms] [k] queue_work_on
> + 0.46% fb.sh [kernel.kallsyms] [k] shrink_slab_memcg
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112561772.4097.11011071937553113003.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063070859.1818.11870882950920963480.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:25 +00:00
|
|
|
*
|
|
|
|
* list_lru_add() shrink_slab_memcg()
|
|
|
|
* list_add_tail() clear_bit()
|
|
|
|
* <MB> <MB>
|
|
|
|
* set_bit() do_shrink_slab()
|
|
|
|
*/
|
|
|
|
smp_mb__after_atomic();
|
|
|
|
ret = do_shrink_slab(&sc, shrinker, priority);
|
|
|
|
if (ret == SHRINK_EMPTY)
|
|
|
|
ret = 0;
|
|
|
|
else
|
2021-05-05 01:36:11 +00:00
|
|
|
set_shrinker_bit(memcg, nid, i);
|
mm/vmscan.c: clear shrinker bit if there are no objects related to memcg
To avoid further unneed calls of do_shrink_slab() for shrinkers, which
already do not have any charged objects in a memcg, their bits have to
be cleared.
This patch introduces a lockless mechanism to do that without races
without parallel list lru add. After do_shrink_slab() returns
SHRINK_EMPTY the first time, we clear the bit and call it once again.
Then we restore the bit, if the new return value is different.
Note, that single smp_mb__after_atomic() in shrink_slab_memcg() covers
two situations:
1)list_lru_add() shrink_slab_memcg
list_add_tail() for_each_set_bit() <--- read bit
do_shrink_slab() <--- missed list update (no barrier)
<MB> <MB>
set_bit() do_shrink_slab() <--- seen list update
This situation, when the first do_shrink_slab() sees set bit, but it
doesn't see list update (i.e., race with the first element queueing), is
rare. So we don't add <MB> before the first call of do_shrink_slab()
instead of this to do not slow down generic case. Also, it's need the
second call as seen in below in (2).
2)list_lru_add() shrink_slab_memcg()
list_add_tail() ...
set_bit() ...
... for_each_set_bit()
do_shrink_slab() do_shrink_slab()
clear_bit() ...
... ...
list_lru_add() ...
list_add_tail() clear_bit()
<MB> <MB>
set_bit() do_shrink_slab()
The barriers guarantee that the second do_shrink_slab() in the right
side task sees list update if really cleared the bit. This case is
drawn in the code comment.
[Results/performance of the patchset]
After the whole patchset applied the below test shows signify increase
of performance:
$echo 1 > /sys/fs/cgroup/memory/memory.use_hierarchy
$mkdir /sys/fs/cgroup/memory/ct
$echo 4000M > /sys/fs/cgroup/memory/ct/memory.kmem.limit_in_bytes
$for i in `seq 0 4000`; do mkdir /sys/fs/cgroup/memory/ct/$i;
echo $$ > /sys/fs/cgroup/memory/ct/$i/cgroup.procs;
mkdir -p s/$i; mount -t tmpfs $i s/$i;
touch s/$i/file; done
Then, 5 sequential calls of drop caches:
$time echo 3 > /proc/sys/vm/drop_caches
1)Before:
0.00user 13.78system 0:13.78elapsed 99%CPU
0.00user 5.59system 0:05.60elapsed 99%CPU
0.00user 5.48system 0:05.48elapsed 99%CPU
0.00user 8.35system 0:08.35elapsed 99%CPU
0.00user 8.34system 0:08.35elapsed 99%CPU
2)After
0.00user 1.10system 0:01.10elapsed 99%CPU
0.00user 0.00system 0:00.01elapsed 64%CPU
0.00user 0.01system 0:00.01elapsed 82%CPU
0.00user 0.00system 0:00.01elapsed 64%CPU
0.00user 0.01system 0:00.01elapsed 82%CPU
The results show the performance increases at least in 548 times.
Shakeel Butt tested this patchset with fork-bomb on his configuration:
> I created 255 memcgs, 255 ext4 mounts and made each memcg create a
> file containing few KiBs on corresponding mount. Then in a separate
> memcg of 200 MiB limit ran a fork-bomb.
>
> I ran the "perf record -ag -- sleep 60" and below are the results:
>
> Without the patch series:
> Samples: 4M of event 'cycles', Event count (approx.): 3279403076005
> + 36.40% fb.sh [kernel.kallsyms] [k] shrink_slab
> + 18.97% fb.sh [kernel.kallsyms] [k] list_lru_count_one
> + 6.75% fb.sh [kernel.kallsyms] [k] super_cache_count
> + 0.49% fb.sh [kernel.kallsyms] [k] down_read_trylock
> + 0.44% fb.sh [kernel.kallsyms] [k] mem_cgroup_iter
> + 0.27% fb.sh [kernel.kallsyms] [k] up_read
> + 0.21% fb.sh [kernel.kallsyms] [k] osq_lock
> + 0.13% fb.sh [kernel.kallsyms] [k] shmem_unused_huge_count
> + 0.08% fb.sh [kernel.kallsyms] [k] shrink_node_memcg
> + 0.08% fb.sh [kernel.kallsyms] [k] shrink_node
>
> With the patch series:
> Samples: 4M of event 'cycles', Event count (approx.): 2756866824946
> + 47.49% fb.sh [kernel.kallsyms] [k] down_read_trylock
> + 30.72% fb.sh [kernel.kallsyms] [k] up_read
> + 9.51% fb.sh [kernel.kallsyms] [k] mem_cgroup_iter
> + 1.69% fb.sh [kernel.kallsyms] [k] shrink_node_memcg
> + 1.35% fb.sh [kernel.kallsyms] [k] mem_cgroup_protected
> + 1.05% fb.sh [kernel.kallsyms] [k] queued_spin_lock_slowpath
> + 0.85% fb.sh [kernel.kallsyms] [k] _raw_spin_lock
> + 0.78% fb.sh [kernel.kallsyms] [k] lruvec_lru_size
> + 0.57% fb.sh [kernel.kallsyms] [k] shrink_node
> + 0.54% fb.sh [kernel.kallsyms] [k] queue_work_on
> + 0.46% fb.sh [kernel.kallsyms] [k] shrink_slab_memcg
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112561772.4097.11011071937553113003.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063070859.1818.11870882950920963480.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:25 +00:00
|
|
|
}
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
freed += ret;
|
|
|
|
|
|
|
|
if (rwsem_is_contended(&shrinker_rwsem)) {
|
|
|
|
freed = freed ? : 1;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
unlock:
|
|
|
|
up_read(&shrinker_rwsem);
|
|
|
|
return freed;
|
|
|
|
}
|
2019-09-23 22:38:12 +00:00
|
|
|
#else /* CONFIG_MEMCG */
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
|
|
|
|
struct mem_cgroup *memcg, int priority)
|
|
|
|
{
|
|
|
|
return 0;
|
|
|
|
}
|
2019-09-23 22:38:12 +00:00
|
|
|
#endif /* CONFIG_MEMCG */
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
/**
|
2015-02-12 22:58:54 +00:00
|
|
|
* shrink_slab - shrink slab caches
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
* @gfp_mask: allocation context
|
|
|
|
* @nid: node whose slab caches to target
|
2015-02-12 22:58:54 +00:00
|
|
|
* @memcg: memory cgroup whose slab caches to target
|
mm: use sc->priority for slab shrink targets
Previously we were using the ratio of the number of lru pages scanned to
the number of eligible lru pages to determine the number of slab objects
to scan. The problem with this is that these two things have nothing to
do with each other, so in slab heavy work loads where there is little to
no page cache we can end up with the pages scanned being a very low
number. This means that we reclaim next to no slab pages and waste a
lot of time reclaiming small amounts of space.
Consider the following scenario, where we have the following values and
the rest of the memory usage is in slab
Active: 58840 kB
Inactive: 46860 kB
Every time we do a get_scan_count() we do this
scan = size >> sc->priority
where sc->priority starts at DEF_PRIORITY, which is 12. The first loop
through reclaim would result in a scan target of 2 pages to 11715 total
inactive pages, and 3 pages to 14710 total active pages. This is a
really really small target for a system that is entirely slab pages.
And this is super optimistic, this assumes we even get to scan these
pages. We don't increment sc->nr_scanned unless we 1) isolate the page,
which assumes it's not in use, and 2) can lock the page. Under pressure
these numbers could probably go down, I'm sure there's some random pages
from daemons that aren't actually in use, so the targets get even
smaller.
Instead use sc->priority in the same way we use it to determine scan
amounts for the lru's. This generally equates to pages. Consider the
following
slab_pages = (nr_objects * object_size) / PAGE_SIZE
What we would like to do is
scan = slab_pages >> sc->priority
but we don't know the number of slab pages each shrinker controls, only
the objects. However say that theoretically we knew how many pages a
shrinker controlled, we'd still have to convert this to objects, which
would look like the following
scan = shrinker_pages >> sc->priority
scan_objects = (PAGE_SIZE / object_size) * scan
or written another way
scan_objects = (shrinker_pages >> sc->priority) *
(PAGE_SIZE / object_size)
which can thus be written
scan_objects = ((shrinker_pages * PAGE_SIZE) / object_size) >>
sc->priority
which is just
scan_objects = nr_objects >> sc->priority
We don't need to know exactly how many pages each shrinker represents,
it's objects are all the information we need. Making this change allows
us to place an appropriate amount of pressure on the shrinker pools for
their relative size.
Link: http://lkml.kernel.org/r/1510780549-6812-1-git-send-email-josef@toxicpanda.com
Signed-off-by: Josef Bacik <jbacik@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Dave Chinner <david@fromorbit.com>
Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-02-01 00:16:26 +00:00
|
|
|
* @priority: the reclaim priority
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
* Call the shrink functions to age shrinkable caches.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
* @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
|
|
|
|
* unaware shrinkers will receive a node id of 0 instead.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2018-08-17 22:48:17 +00:00
|
|
|
* @memcg specifies the memory cgroup to target. Unaware shrinkers
|
|
|
|
* are called only if it is the root cgroup.
|
2015-02-12 22:58:54 +00:00
|
|
|
*
|
mm: use sc->priority for slab shrink targets
Previously we were using the ratio of the number of lru pages scanned to
the number of eligible lru pages to determine the number of slab objects
to scan. The problem with this is that these two things have nothing to
do with each other, so in slab heavy work loads where there is little to
no page cache we can end up with the pages scanned being a very low
number. This means that we reclaim next to no slab pages and waste a
lot of time reclaiming small amounts of space.
Consider the following scenario, where we have the following values and
the rest of the memory usage is in slab
Active: 58840 kB
Inactive: 46860 kB
Every time we do a get_scan_count() we do this
scan = size >> sc->priority
where sc->priority starts at DEF_PRIORITY, which is 12. The first loop
through reclaim would result in a scan target of 2 pages to 11715 total
inactive pages, and 3 pages to 14710 total active pages. This is a
really really small target for a system that is entirely slab pages.
And this is super optimistic, this assumes we even get to scan these
pages. We don't increment sc->nr_scanned unless we 1) isolate the page,
which assumes it's not in use, and 2) can lock the page. Under pressure
these numbers could probably go down, I'm sure there's some random pages
from daemons that aren't actually in use, so the targets get even
smaller.
Instead use sc->priority in the same way we use it to determine scan
amounts for the lru's. This generally equates to pages. Consider the
following
slab_pages = (nr_objects * object_size) / PAGE_SIZE
What we would like to do is
scan = slab_pages >> sc->priority
but we don't know the number of slab pages each shrinker controls, only
the objects. However say that theoretically we knew how many pages a
shrinker controlled, we'd still have to convert this to objects, which
would look like the following
scan = shrinker_pages >> sc->priority
scan_objects = (PAGE_SIZE / object_size) * scan
or written another way
scan_objects = (shrinker_pages >> sc->priority) *
(PAGE_SIZE / object_size)
which can thus be written
scan_objects = ((shrinker_pages * PAGE_SIZE) / object_size) >>
sc->priority
which is just
scan_objects = nr_objects >> sc->priority
We don't need to know exactly how many pages each shrinker represents,
it's objects are all the information we need. Making this change allows
us to place an appropriate amount of pressure on the shrinker pools for
their relative size.
Link: http://lkml.kernel.org/r/1510780549-6812-1-git-send-email-josef@toxicpanda.com
Signed-off-by: Josef Bacik <jbacik@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Dave Chinner <david@fromorbit.com>
Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-02-01 00:16:26 +00:00
|
|
|
* @priority is sc->priority, we take the number of objects and >> by priority
|
|
|
|
* in order to get the scan target.
|
2005-06-22 00:14:35 +00:00
|
|
|
*
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
* Returns the number of reclaimed slab objects.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2015-02-12 22:58:54 +00:00
|
|
|
static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
|
|
|
|
struct mem_cgroup *memcg,
|
mm: use sc->priority for slab shrink targets
Previously we were using the ratio of the number of lru pages scanned to
the number of eligible lru pages to determine the number of slab objects
to scan. The problem with this is that these two things have nothing to
do with each other, so in slab heavy work loads where there is little to
no page cache we can end up with the pages scanned being a very low
number. This means that we reclaim next to no slab pages and waste a
lot of time reclaiming small amounts of space.
Consider the following scenario, where we have the following values and
the rest of the memory usage is in slab
Active: 58840 kB
Inactive: 46860 kB
Every time we do a get_scan_count() we do this
scan = size >> sc->priority
where sc->priority starts at DEF_PRIORITY, which is 12. The first loop
through reclaim would result in a scan target of 2 pages to 11715 total
inactive pages, and 3 pages to 14710 total active pages. This is a
really really small target for a system that is entirely slab pages.
And this is super optimistic, this assumes we even get to scan these
pages. We don't increment sc->nr_scanned unless we 1) isolate the page,
which assumes it's not in use, and 2) can lock the page. Under pressure
these numbers could probably go down, I'm sure there's some random pages
from daemons that aren't actually in use, so the targets get even
smaller.
Instead use sc->priority in the same way we use it to determine scan
amounts for the lru's. This generally equates to pages. Consider the
following
slab_pages = (nr_objects * object_size) / PAGE_SIZE
What we would like to do is
scan = slab_pages >> sc->priority
but we don't know the number of slab pages each shrinker controls, only
the objects. However say that theoretically we knew how many pages a
shrinker controlled, we'd still have to convert this to objects, which
would look like the following
scan = shrinker_pages >> sc->priority
scan_objects = (PAGE_SIZE / object_size) * scan
or written another way
scan_objects = (shrinker_pages >> sc->priority) *
(PAGE_SIZE / object_size)
which can thus be written
scan_objects = ((shrinker_pages * PAGE_SIZE) / object_size) >>
sc->priority
which is just
scan_objects = nr_objects >> sc->priority
We don't need to know exactly how many pages each shrinker represents,
it's objects are all the information we need. Making this change allows
us to place an appropriate amount of pressure on the shrinker pools for
their relative size.
Link: http://lkml.kernel.org/r/1510780549-6812-1-git-send-email-josef@toxicpanda.com
Signed-off-by: Josef Bacik <jbacik@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Dave Chinner <david@fromorbit.com>
Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-02-01 00:16:26 +00:00
|
|
|
int priority)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2018-10-05 22:52:10 +00:00
|
|
|
unsigned long ret, freed = 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
struct shrinker *shrinker;
|
|
|
|
|
2019-08-03 04:48:44 +00:00
|
|
|
/*
|
|
|
|
* The root memcg might be allocated even though memcg is disabled
|
|
|
|
* via "cgroup_disable=memory" boot parameter. This could make
|
|
|
|
* mem_cgroup_is_root() return false, then just run memcg slab
|
|
|
|
* shrink, but skip global shrink. This may result in premature
|
|
|
|
* oom.
|
|
|
|
*/
|
|
|
|
if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
|
mm/vmscan.c: iterate only over charged shrinkers during memcg shrink_slab()
Using the preparations made in previous patches, in case of memcg
shrink, we may avoid shrinkers, which are not set in memcg's shrinkers
bitmap. To do that, we separate iterations over memcg-aware and
!memcg-aware shrinkers, and memcg-aware shrinkers are chosen via
for_each_set_bit() from the bitmap. In case of big nodes, having many
isolated environments, this gives significant performance growth. See
next patches for the details.
Note that the patch does not respect to empty memcg shrinkers, since we
never clear the bitmap bits after we set it once. Their shrinkers will
be called again, with no shrinked objects as result. This functionality
is provided by next patches.
[ktkhai@virtuozzo.com: v9]
Link: http://lkml.kernel.org/r/153112558507.4097.12713813335683345488.stgit@localhost.localdomain
Link: http://lkml.kernel.org/r/153063066653.1818.976035462801487910.stgit@localhost.localdomain
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Acked-by: Vladimir Davydov <vdavydov.dev@gmail.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Li RongQing <lirongqing@baidu.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Matthias Kaehlcke <mka@chromium.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Philippe Ombredanne <pombredanne@nexb.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Sahitya Tummala <stummala@codeaurora.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <longman@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 22:48:14 +00:00
|
|
|
return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
|
2015-02-12 22:58:54 +00:00
|
|
|
|
2018-04-05 23:23:35 +00:00
|
|
|
if (!down_read_trylock(&shrinker_rwsem))
|
2011-05-25 00:11:11 +00:00
|
|
|
goto out;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
list_for_each_entry(shrinker, &shrinker_list, list) {
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
struct shrink_control sc = {
|
|
|
|
.gfp_mask = gfp_mask,
|
|
|
|
.nid = nid,
|
2015-02-12 22:58:54 +00:00
|
|
|
.memcg = memcg,
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
};
|
2014-01-23 23:53:23 +00:00
|
|
|
|
2018-08-17 22:48:21 +00:00
|
|
|
ret = do_shrink_slab(&sc, shrinker, priority);
|
|
|
|
if (ret == SHRINK_EMPTY)
|
|
|
|
ret = 0;
|
|
|
|
freed += ret;
|
2018-02-01 00:16:55 +00:00
|
|
|
/*
|
|
|
|
* Bail out if someone want to register a new shrinker to
|
2020-06-04 23:49:10 +00:00
|
|
|
* prevent the registration from being stalled for long periods
|
2018-02-01 00:16:55 +00:00
|
|
|
* by parallel ongoing shrinking.
|
|
|
|
*/
|
|
|
|
if (rwsem_is_contended(&shrinker_rwsem)) {
|
|
|
|
freed = freed ? : 1;
|
|
|
|
break;
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
up_read(&shrinker_rwsem);
|
2011-05-25 00:11:11 +00:00
|
|
|
out:
|
|
|
|
cond_resched();
|
2013-08-28 00:17:56 +00:00
|
|
|
return freed;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2022-11-15 12:32:55 +00:00
|
|
|
static unsigned long drop_slab_node(int nid)
|
2015-02-12 22:58:54 +00:00
|
|
|
{
|
2022-11-15 12:32:55 +00:00
|
|
|
unsigned long freed = 0;
|
|
|
|
struct mem_cgroup *memcg = NULL;
|
2015-02-12 22:58:54 +00:00
|
|
|
|
2022-11-15 12:32:55 +00:00
|
|
|
memcg = mem_cgroup_iter(NULL, NULL, NULL);
|
2015-02-12 22:58:54 +00:00
|
|
|
do {
|
2022-11-15 12:32:55 +00:00
|
|
|
freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
|
|
|
|
} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
|
mm/vmscan: fix infinite loop in drop_slab_node
We have observed that drop_caches can take a considerable amount of
time (<put data here>). Especially when there are many memcgs involved
because they are adding an additional overhead.
It is quite unfortunate that the operation cannot be interrupted by a
signal currently. Add a check for fatal signals into the main loop so
that userspace can control early bailout.
There are two reasons:
1. We have too many memcgs, even though one object freed in one memcg,
the sum of object is bigger than 10.
2. We spend a lot of time in traverse memcg once. So, the memcg who
traversed at the first have been freed many objects. Traverse memcg
next time, the freed count bigger than 10 again.
We can get the following info through 'ps':
root:~# ps -aux | grep drop
root 357956 ... R Aug25 21119854:55 echo 3 > /proc/sys/vm/drop_caches
root 1771385 ... R Aug16 21146421:17 echo 3 > /proc/sys/vm/drop_caches
root 1986319 ... R 18:56 117:27 echo 3 > /proc/sys/vm/drop_caches
root 2002148 ... R Aug24 5720:39 echo 3 > /proc/sys/vm/drop_caches
root 2564666 ... R 18:59 113:58 echo 3 > /proc/sys/vm/drop_caches
root 2639347 ... R Sep03 2383:39 echo 3 > /proc/sys/vm/drop_caches
root 3904747 ... R 03:35 993:31 echo 3 > /proc/sys/vm/drop_caches
root 4016780 ... R Aug21 7882:18 echo 3 > /proc/sys/vm/drop_caches
Use bpftrace follow 'freed' value in drop_slab_node:
root:~# bpftrace -e 'kprobe:drop_slab_node+70 {@ret=hist(reg("bp")); }'
Attaching 1 probe...
^B^C
@ret:
[64, 128) 1 | |
[128, 256) 28 | |
[256, 512) 107 |@ |
[512, 1K) 298 |@@@ |
[1K, 2K) 613 |@@@@@@@ |
[2K, 4K) 4435 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[4K, 8K) 442 |@@@@@ |
[8K, 16K) 299 |@@@ |
[16K, 32K) 100 |@ |
[32K, 64K) 139 |@ |
[64K, 128K) 56 | |
[128K, 256K) 26 | |
[256K, 512K) 2 | |
In the while loop, we can check whether the TASK_KILLABLE signal is set,
if so, we should break the loop.
Signed-off-by: Chunxin Zang <zangchunxin@bytedance.com>
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Chris Down <chris@chrisdown.name>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Matthew Wilcox <willy@infradead.org>
Link: https://lkml.kernel.org/r/20200909152047.27905-1-zangchunxin@bytedance.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-13 23:56:46 +00:00
|
|
|
|
2022-11-15 12:32:55 +00:00
|
|
|
return freed;
|
2015-02-12 22:58:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void drop_slab(void)
|
|
|
|
{
|
|
|
|
int nid;
|
2022-11-15 12:32:55 +00:00
|
|
|
int shift = 0;
|
|
|
|
unsigned long freed;
|
|
|
|
|
|
|
|
do {
|
|
|
|
freed = 0;
|
|
|
|
for_each_online_node(nid) {
|
|
|
|
if (fatal_signal_pending(current))
|
|
|
|
return;
|
2015-02-12 22:58:54 +00:00
|
|
|
|
2022-11-15 12:32:55 +00:00
|
|
|
freed += drop_slab_node(nid);
|
|
|
|
}
|
|
|
|
} while ((freed >> shift++) > 1);
|
2015-02-12 22:58:54 +00:00
|
|
|
}
|
|
|
|
|
2022-10-26 18:01:33 +00:00
|
|
|
static int reclaimer_offset(void)
|
|
|
|
{
|
|
|
|
BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
|
|
|
|
PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD);
|
|
|
|
BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD !=
|
|
|
|
PGSCAN_DIRECT - PGSCAN_KSWAPD);
|
|
|
|
BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
|
|
|
|
PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD);
|
|
|
|
BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD !=
|
|
|
|
PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD);
|
|
|
|
|
|
|
|
if (current_is_kswapd())
|
|
|
|
return 0;
|
|
|
|
if (current_is_khugepaged())
|
|
|
|
return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD;
|
|
|
|
return PGSTEAL_DIRECT - PGSTEAL_KSWAPD;
|
|
|
|
}
|
|
|
|
|
2022-01-18 04:35:57 +00:00
|
|
|
static inline int is_page_cache_freeable(struct folio *folio)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2009-09-22 00:03:00 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* A freeable page cache folio is referenced only by the caller
|
|
|
|
* that isolated the folio, the page cache and optional filesystem
|
|
|
|
* private data at folio->private.
|
2009-09-22 00:03:00 +00:00
|
|
|
*/
|
2022-01-18 04:35:57 +00:00
|
|
|
return folio_ref_count(folio) - folio_test_private(folio) ==
|
|
|
|
1 + folio_nr_pages(folio);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2022-01-18 04:35:57 +00:00
|
|
|
* We detected a synchronous write error writing a folio out. Probably
|
2005-04-16 22:20:36 +00:00
|
|
|
* -ENOSPC. We need to propagate that into the address_space for a subsequent
|
|
|
|
* fsync(), msync() or close().
|
|
|
|
*
|
|
|
|
* The tricky part is that after writepage we cannot touch the mapping: nothing
|
2022-01-18 04:35:57 +00:00
|
|
|
* prevents it from being freed up. But we have a ref on the folio and once
|
|
|
|
* that folio is locked, the mapping is pinned.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2022-01-18 04:35:57 +00:00
|
|
|
* We're allowed to run sleeping folio_lock() here because we know the caller has
|
2005-04-16 22:20:36 +00:00
|
|
|
* __GFP_FS.
|
|
|
|
*/
|
|
|
|
static void handle_write_error(struct address_space *mapping,
|
2022-01-18 04:35:57 +00:00
|
|
|
struct folio *folio, int error)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2022-01-18 04:35:57 +00:00
|
|
|
folio_lock(folio);
|
|
|
|
if (folio_mapping(folio) == mapping)
|
2007-05-08 07:23:25 +00:00
|
|
|
mapping_set_error(mapping, error);
|
2022-01-18 04:35:57 +00:00
|
|
|
folio_unlock(folio);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
static bool skip_throttle_noprogress(pg_data_t *pgdat)
|
|
|
|
{
|
|
|
|
int reclaimable = 0, write_pending = 0;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If kswapd is disabled, reschedule if necessary but do not
|
|
|
|
* throttle as the system is likely near OOM.
|
|
|
|
*/
|
|
|
|
if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* If there are a lot of dirty/writeback folios then do not
|
|
|
|
* throttle as throttling will occur when the folios cycle
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
* towards the end of the LRU if still under writeback.
|
|
|
|
*/
|
|
|
|
for (i = 0; i < MAX_NR_ZONES; i++) {
|
|
|
|
struct zone *zone = pgdat->node_zones + i;
|
|
|
|
|
2022-04-29 06:16:03 +00:00
|
|
|
if (!managed_zone(zone))
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
reclaimable += zone_reclaimable_pages(zone);
|
|
|
|
write_pending += zone_page_state_snapshot(zone,
|
|
|
|
NR_ZONE_WRITE_PENDING);
|
|
|
|
}
|
|
|
|
if (2 * write_pending <= reclaimable)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2021-11-05 20:42:42 +00:00
|
|
|
void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
{
|
|
|
|
wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
|
2021-11-05 20:42:42 +00:00
|
|
|
long timeout, ret;
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
DEFINE_WAIT(wait);
|
|
|
|
|
|
|
|
/*
|
2023-03-10 22:03:26 +00:00
|
|
|
* Do not throttle user workers, kthreads other than kswapd or
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
* workqueues. They may be required for reclaim to make
|
|
|
|
* forward progress (e.g. journalling workqueues or kthreads).
|
|
|
|
*/
|
|
|
|
if (!current_is_kswapd() &&
|
2023-03-10 22:03:26 +00:00
|
|
|
current->flags & (PF_USER_WORKER|PF_KTHREAD)) {
|
2022-02-12 00:32:29 +00:00
|
|
|
cond_resched();
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
return;
|
2022-02-12 00:32:29 +00:00
|
|
|
}
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
|
2021-11-05 20:42:42 +00:00
|
|
|
/*
|
|
|
|
* These figures are pulled out of thin air.
|
|
|
|
* VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
|
|
|
|
* parallel reclaimers which is a short-lived event so the timeout is
|
|
|
|
* short. Failing to make progress or waiting on writeback are
|
|
|
|
* potentially long-lived events so use a longer timeout. This is shaky
|
|
|
|
* logic as a failure to make progress could be due to anything from
|
2022-09-02 19:45:57 +00:00
|
|
|
* writeback to a slow device to excessive referenced folios at the tail
|
2021-11-05 20:42:42 +00:00
|
|
|
* of the inactive LRU.
|
|
|
|
*/
|
|
|
|
switch(reason) {
|
|
|
|
case VMSCAN_THROTTLE_WRITEBACK:
|
|
|
|
timeout = HZ/10;
|
|
|
|
|
|
|
|
if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
|
|
|
|
WRITE_ONCE(pgdat->nr_reclaim_start,
|
|
|
|
node_page_state(pgdat, NR_THROTTLED_WRITTEN));
|
|
|
|
}
|
|
|
|
|
|
|
|
break;
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
case VMSCAN_THROTTLE_CONGESTED:
|
|
|
|
fallthrough;
|
2021-11-05 20:42:42 +00:00
|
|
|
case VMSCAN_THROTTLE_NOPROGRESS:
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
if (skip_throttle_noprogress(pgdat)) {
|
|
|
|
cond_resched();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
timeout = 1;
|
|
|
|
|
2021-11-05 20:42:42 +00:00
|
|
|
break;
|
|
|
|
case VMSCAN_THROTTLE_ISOLATED:
|
|
|
|
timeout = HZ/50;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
WARN_ON_ONCE(1);
|
|
|
|
timeout = HZ;
|
|
|
|
break;
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
|
|
|
|
ret = schedule_timeout(timeout);
|
|
|
|
finish_wait(wqh, &wait);
|
2021-11-05 20:42:29 +00:00
|
|
|
|
2021-11-05 20:42:42 +00:00
|
|
|
if (reason == VMSCAN_THROTTLE_WRITEBACK)
|
2021-11-05 20:42:29 +00:00
|
|
|
atomic_dec(&pgdat->nr_writeback_throttled);
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
|
|
|
|
trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
|
|
|
|
jiffies_to_usecs(timeout - ret),
|
|
|
|
reason);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* Account for folios written if tasks are throttled waiting on dirty
|
|
|
|
* folios to clean. If enough folios have been cleaned since throttling
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
* started then wakeup the throttled tasks.
|
|
|
|
*/
|
Merge branch 'akpm' (patches from Andrew)
Merge misc updates from Andrew Morton:
"257 patches.
Subsystems affected by this patch series: scripts, ocfs2, vfs, and
mm (slab-generic, slab, slub, kconfig, dax, kasan, debug, pagecache,
gup, swap, memcg, pagemap, mprotect, mremap, iomap, tracing, vmalloc,
pagealloc, memory-failure, hugetlb, userfaultfd, vmscan, tools,
memblock, oom-kill, hugetlbfs, migration, thp, readahead, nommu, ksm,
vmstat, madvise, memory-hotplug, rmap, zsmalloc, highmem, zram,
cleanups, kfence, and damon)"
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (257 commits)
mm/damon: remove return value from before_terminate callback
mm/damon: fix a few spelling mistakes in comments and a pr_debug message
mm/damon: simplify stop mechanism
Docs/admin-guide/mm/pagemap: wordsmith page flags descriptions
Docs/admin-guide/mm/damon/start: simplify the content
Docs/admin-guide/mm/damon/start: fix a wrong link
Docs/admin-guide/mm/damon/start: fix wrong example commands
mm/damon/dbgfs: add adaptive_targets list check before enable monitor_on
mm/damon: remove unnecessary variable initialization
Documentation/admin-guide/mm/damon: add a document for DAMON_RECLAIM
mm/damon: introduce DAMON-based Reclamation (DAMON_RECLAIM)
selftests/damon: support watermarks
mm/damon/dbgfs: support watermarks
mm/damon/schemes: activate schemes based on a watermarks mechanism
tools/selftests/damon: update for regions prioritization of schemes
mm/damon/dbgfs: support prioritization weights
mm/damon/vaddr,paddr: support pageout prioritization
mm/damon/schemes: prioritize regions within the quotas
mm/damon/selftests: support schemes quotas
mm/damon/dbgfs: support quotas of schemes
...
2021-11-06 21:08:17 +00:00
|
|
|
void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
int nr_throttled)
|
|
|
|
{
|
|
|
|
unsigned long nr_written;
|
|
|
|
|
Merge branch 'akpm' (patches from Andrew)
Merge misc updates from Andrew Morton:
"257 patches.
Subsystems affected by this patch series: scripts, ocfs2, vfs, and
mm (slab-generic, slab, slub, kconfig, dax, kasan, debug, pagecache,
gup, swap, memcg, pagemap, mprotect, mremap, iomap, tracing, vmalloc,
pagealloc, memory-failure, hugetlb, userfaultfd, vmscan, tools,
memblock, oom-kill, hugetlbfs, migration, thp, readahead, nommu, ksm,
vmstat, madvise, memory-hotplug, rmap, zsmalloc, highmem, zram,
cleanups, kfence, and damon)"
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (257 commits)
mm/damon: remove return value from before_terminate callback
mm/damon: fix a few spelling mistakes in comments and a pr_debug message
mm/damon: simplify stop mechanism
Docs/admin-guide/mm/pagemap: wordsmith page flags descriptions
Docs/admin-guide/mm/damon/start: simplify the content
Docs/admin-guide/mm/damon/start: fix a wrong link
Docs/admin-guide/mm/damon/start: fix wrong example commands
mm/damon/dbgfs: add adaptive_targets list check before enable monitor_on
mm/damon: remove unnecessary variable initialization
Documentation/admin-guide/mm/damon: add a document for DAMON_RECLAIM
mm/damon: introduce DAMON-based Reclamation (DAMON_RECLAIM)
selftests/damon: support watermarks
mm/damon/dbgfs: support watermarks
mm/damon/schemes: activate schemes based on a watermarks mechanism
tools/selftests/damon: update for regions prioritization of schemes
mm/damon/dbgfs: support prioritization weights
mm/damon/vaddr,paddr: support pageout prioritization
mm/damon/schemes: prioritize regions within the quotas
mm/damon/selftests: support schemes quotas
mm/damon/dbgfs: support quotas of schemes
...
2021-11-06 21:08:17 +00:00
|
|
|
node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* This is an inaccurate read as the per-cpu deltas may not
|
|
|
|
* be synchronised. However, given that the system is
|
|
|
|
* writeback throttled, it is not worth taking the penalty
|
|
|
|
* of getting an accurate count. At worst, the throttle
|
|
|
|
* timeout guarantees forward progress.
|
|
|
|
*/
|
|
|
|
nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
|
|
|
|
READ_ONCE(pgdat->nr_reclaim_start);
|
|
|
|
|
|
|
|
if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
|
|
|
|
wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
|
|
|
|
}
|
|
|
|
|
2006-06-23 09:03:38 +00:00
|
|
|
/* possible outcome of pageout() */
|
|
|
|
typedef enum {
|
2022-09-02 19:45:57 +00:00
|
|
|
/* failed to write folio out, folio is locked */
|
2006-06-23 09:03:38 +00:00
|
|
|
PAGE_KEEP,
|
2022-09-02 19:45:57 +00:00
|
|
|
/* move folio to the active list, folio is locked */
|
2006-06-23 09:03:38 +00:00
|
|
|
PAGE_ACTIVATE,
|
2022-09-02 19:45:57 +00:00
|
|
|
/* folio has been sent to the disk successfully, folio is unlocked */
|
2006-06-23 09:03:38 +00:00
|
|
|
PAGE_SUCCESS,
|
2022-09-02 19:45:57 +00:00
|
|
|
/* folio is clean and locked */
|
2006-06-23 09:03:38 +00:00
|
|
|
PAGE_CLEAN,
|
|
|
|
} pageout_t;
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* pageout is called by shrink_folio_list() for each dirty folio.
|
[PATCH] vmscan: rename functions
We have:
try_to_free_pages
->shrink_caches(struct zone **zones, ..)
->shrink_zone(struct zone *, ...)
->shrink_cache(struct zone *, ...)
->shrink_list(struct list_head *, ...)
->refill_inactive_list((struct zone *, ...)
which is fairly irrational.
Rename things so that we have
try_to_free_pages
->shrink_zones(struct zone **zones, ..)
->shrink_zone(struct zone *, ...)
->shrink_inactive_list(struct zone *, ...)
->shrink_page_list(struct list_head *, ...)
->shrink_active_list(struct zone *, ...)
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Christoph Lameter <christoph@lameter.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-22 08:08:21 +00:00
|
|
|
* Calls ->writepage().
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2022-05-10 01:20:49 +00:00
|
|
|
static pageout_t pageout(struct folio *folio, struct address_space *mapping,
|
|
|
|
struct swap_iocb **plug)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
|
|
|
/*
|
2022-01-18 04:35:57 +00:00
|
|
|
* If the folio is dirty, only perform writeback if that write
|
2005-04-16 22:20:36 +00:00
|
|
|
* will be non-blocking. To prevent this allocation from being
|
|
|
|
* stalled by pagecache activity. But note that there may be
|
|
|
|
* stalls if we need to run get_block(). We could test
|
|
|
|
* PagePrivate for that.
|
|
|
|
*
|
2014-04-03 07:17:43 +00:00
|
|
|
* If this process is currently in __generic_file_write_iter() against
|
2022-01-18 04:35:57 +00:00
|
|
|
* this folio's queue, we can perform writeback even if that
|
2005-04-16 22:20:36 +00:00
|
|
|
* will block.
|
|
|
|
*
|
2022-01-18 04:35:57 +00:00
|
|
|
* If the folio is swapcache, write it back even if that would
|
2005-04-16 22:20:36 +00:00
|
|
|
* block, for some throttling. This happens by accident, because
|
|
|
|
* swap_backing_dev_info is bust: it doesn't reflect the
|
|
|
|
* congestion state of the swapdevs. Easy to fix, if needed.
|
|
|
|
*/
|
2022-01-18 04:35:57 +00:00
|
|
|
if (!is_page_cache_freeable(folio))
|
2005-04-16 22:20:36 +00:00
|
|
|
return PAGE_KEEP;
|
|
|
|
if (!mapping) {
|
|
|
|
/*
|
2022-01-18 04:35:57 +00:00
|
|
|
* Some data journaling orphaned folios can have
|
|
|
|
* folio->mapping == NULL while being dirty with clean buffers.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2022-01-18 04:35:57 +00:00
|
|
|
if (folio_test_private(folio)) {
|
2022-05-01 05:08:08 +00:00
|
|
|
if (try_to_free_buffers(folio)) {
|
2022-01-18 04:35:57 +00:00
|
|
|
folio_clear_dirty(folio);
|
|
|
|
pr_info("%s: orphaned folio\n", __func__);
|
2005-04-16 22:20:36 +00:00
|
|
|
return PAGE_CLEAN;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return PAGE_KEEP;
|
|
|
|
}
|
|
|
|
if (mapping->a_ops->writepage == NULL)
|
|
|
|
return PAGE_ACTIVATE;
|
|
|
|
|
2022-01-18 04:35:57 +00:00
|
|
|
if (folio_clear_dirty_for_io(folio)) {
|
2005-04-16 22:20:36 +00:00
|
|
|
int res;
|
|
|
|
struct writeback_control wbc = {
|
|
|
|
.sync_mode = WB_SYNC_NONE,
|
|
|
|
.nr_to_write = SWAP_CLUSTER_MAX,
|
[PATCH] writeback: fix range handling
When a writeback_control's `start' and `end' fields are used to
indicate a one-byte-range starting at file offset zero, the required
values of .start=0,.end=0 mean that the ->writepages() implementation
has no way of telling that it is being asked to perform a range
request. Because we're currently overloading (start == 0 && end == 0)
to mean "this is not a write-a-range request".
To make all this sane, the patch changes range of writeback_control.
So caller does: If it is calling ->writepages() to write pages, it
sets range (range_start/end or range_cyclic) always.
And if range_cyclic is true, ->writepages() thinks the range is
cyclic, otherwise it just uses range_start and range_end.
This patch does,
- Add LLONG_MAX, LLONG_MIN, ULLONG_MAX to include/linux/kernel.h
-1 is usually ok for range_end (type is long long). But, if someone did,
range_end += val; range_end is "val - 1"
u64val = range_end >> bits; u64val is "~(0ULL)"
or something, they are wrong. So, this adds LLONG_MAX to avoid nasty
things, and uses LLONG_MAX for range_end.
- All callers of ->writepages() sets range_start/end or range_cyclic.
- Fix updates of ->writeback_index. It seems already bit strange.
If it starts at 0 and ended by check of nr_to_write, this last
index may reduce chance to scan end of file. So, this updates
->writeback_index only if range_cyclic is true or whole-file is
scanned.
Signed-off-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Cc: Nathan Scott <nathans@sgi.com>
Cc: Anton Altaparmakov <aia21@cantab.net>
Cc: Steven French <sfrench@us.ibm.com>
Cc: "Vladimir V. Saveliev" <vs@namesys.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-23 09:03:26 +00:00
|
|
|
.range_start = 0,
|
|
|
|
.range_end = LLONG_MAX,
|
2005-04-16 22:20:36 +00:00
|
|
|
.for_reclaim = 1,
|
2022-05-10 01:20:49 +00:00
|
|
|
.swap_plug = plug,
|
2005-04-16 22:20:36 +00:00
|
|
|
};
|
|
|
|
|
2022-01-18 04:35:57 +00:00
|
|
|
folio_set_reclaim(folio);
|
|
|
|
res = mapping->a_ops->writepage(&folio->page, &wbc);
|
2005-04-16 22:20:36 +00:00
|
|
|
if (res < 0)
|
2022-01-18 04:35:57 +00:00
|
|
|
handle_write_error(mapping, folio, res);
|
2005-12-15 22:28:17 +00:00
|
|
|
if (res == AOP_WRITEPAGE_ACTIVATE) {
|
2022-01-18 04:35:57 +00:00
|
|
|
folio_clear_reclaim(folio);
|
2005-04-16 22:20:36 +00:00
|
|
|
return PAGE_ACTIVATE;
|
|
|
|
}
|
2007-08-22 21:01:26 +00:00
|
|
|
|
2022-01-18 04:35:57 +00:00
|
|
|
if (!folio_test_writeback(folio)) {
|
2005-04-16 22:20:36 +00:00
|
|
|
/* synchronous write or broken a_ops? */
|
2022-01-18 04:35:57 +00:00
|
|
|
folio_clear_reclaim(folio);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2022-01-18 04:35:57 +00:00
|
|
|
trace_mm_vmscan_write_folio(folio);
|
|
|
|
node_stat_add_folio(folio, NR_VMSCAN_WRITE);
|
2005-04-16 22:20:36 +00:00
|
|
|
return PAGE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
return PAGE_CLEAN;
|
|
|
|
}
|
|
|
|
|
2006-10-17 07:09:36 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* Same as remove_mapping, but if the folio is removed from the mapping, it
|
2008-07-26 02:45:30 +00:00
|
|
|
* gets returned with a refcount of 0.
|
2006-10-17 07:09:36 +00:00
|
|
|
*/
|
2021-12-23 21:50:23 +00:00
|
|
|
static int __remove_mapping(struct address_space *mapping, struct folio *folio,
|
mm: vmscan: detect file thrashing at the reclaim root
We use refault information to determine whether the cache workingset is
stable or transitioning, and dynamically adjust the inactive:active file
LRU ratio so as to maximize protection from one-off cache during stable
periods, and minimize IO during transitions.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, refaults only affect the
local LRU order in the cgroup in which they are occuring. As a result,
cache transitions can take longer in a cgrouped system as the active pages
of sibling cgroups aren't challenged when they should be.
[ Right now, this is somewhat theoretical, because the siblings, under
continued regular reclaim pressure, should eventually run out of
inactive pages - and since inactive:active *size* balancing is also
done on a cgroup-local level, we will challenge the active pages
eventually in most cases. But the next patch will move that relative
size enforcement to the reclaim root as well, and then this patch
here will be necessary to propagate refault pressure to siblings. ]
This patch moves refault detection to the root of reclaim. Instead of
remembering the cgroup owner of an evicted page, remember the cgroup that
caused the reclaim to happen. When refaults later occur, they'll
correctly influence the cross-cgroup LRU order that reclaim follows.
I.e. if global reclaim kicked out pages in some subgroup A/B/C, the
refault of those pages will challenge the global LRU order, and not just
the local order down inside C.
[hannes@cmpxchg.org: use page_memcg() instead of another lookup]
Link: http://lkml.kernel.org/r/20191115160722.GA309754@cmpxchg.org
Link: http://lkml.kernel.org/r/20191107205334.158354-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:59 +00:00
|
|
|
bool reclaimed, struct mem_cgroup *target_memcg)
|
2006-01-08 09:00:48 +00:00
|
|
|
{
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
int refcount;
|
2020-08-12 01:30:50 +00:00
|
|
|
void *shadow = NULL;
|
memcg: add per cgroup dirty page accounting
When modifying PG_Dirty on cached file pages, update the new
MEM_CGROUP_STAT_DIRTY counter. This is done in the same places where
global NR_FILE_DIRTY is managed. The new memcg stat is visible in the
per memcg memory.stat cgroupfs file. The most recent past attempt at
this was http://thread.gmane.org/gmane.linux.kernel.cgroups/8632
The new accounting supports future efforts to add per cgroup dirty
page throttling and writeback. It also helps an administrator break
down a container's memory usage and provides evidence to understand
memcg oom kills (the new dirty count is included in memcg oom kill
messages).
The ability to move page accounting between memcg
(memory.move_charge_at_immigrate) makes this accounting more
complicated than the global counter. The existing
mem_cgroup_{begin,end}_page_stat() lock is used to serialize move
accounting with stat updates.
Typical update operation:
memcg = mem_cgroup_begin_page_stat(page)
if (TestSetPageDirty()) {
[...]
mem_cgroup_update_page_stat(memcg)
}
mem_cgroup_end_page_stat(memcg)
Summary of mem_cgroup_end_page_stat() overhead:
- Without CONFIG_MEMCG it's a no-op
- With CONFIG_MEMCG and no inter memcg task movement, it's just
rcu_read_lock()
- With CONFIG_MEMCG and inter memcg task movement, it's
rcu_read_lock() + spin_lock_irqsave()
A memcg parameter is added to several routines because their callers
now grab mem_cgroup_begin_page_stat() which returns the memcg later
needed by for mem_cgroup_update_page_stat().
Because mem_cgroup_begin_page_stat() may disable interrupts, some
adjustments are needed:
- move __mark_inode_dirty() from __set_page_dirty() to its caller.
__mark_inode_dirty() locking does not want interrupts disabled.
- use spin_lock_irqsave(tree_lock) rather than spin_lock_irq() in
__delete_from_page_cache(), replace_page_cache_page(),
invalidate_complete_page2(), and __remove_mapping().
text data bss dec hex filename
8925147 1774832 1785856 12485835 be84cb vmlinux-!CONFIG_MEMCG-before
8925339 1774832 1785856 12486027 be858b vmlinux-!CONFIG_MEMCG-after
+192 text bytes
8965977 1784992 1785856 12536825 bf4bf9 vmlinux-CONFIG_MEMCG-before
8966750 1784992 1785856 12537598 bf4efe vmlinux-CONFIG_MEMCG-after
+773 text bytes
Performance tests run on v4.0-rc1-36-g4f671fe2f952. Lower is better for
all metrics, they're all wall clock or cycle counts. The read and write
fault benchmarks just measure fault time, they do not include I/O time.
* CONFIG_MEMCG not set:
baseline patched
kbuild 1m25.030000(+-0.088% 3 samples) 1m25.426667(+-0.120% 3 samples)
dd write 100 MiB 0.859211561 +-15.10% 0.874162885 +-15.03%
dd write 200 MiB 1.670653105 +-17.87% 1.669384764 +-11.99%
dd write 1000 MiB 8.434691190 +-14.15% 8.474733215 +-14.77%
read fault cycles 254.0(+-0.000% 10 samples) 253.0(+-0.000% 10 samples)
write fault cycles 2021.2(+-3.070% 10 samples) 1984.5(+-1.036% 10 samples)
* CONFIG_MEMCG=y root_memcg:
baseline patched
kbuild 1m25.716667(+-0.105% 3 samples) 1m25.686667(+-0.153% 3 samples)
dd write 100 MiB 0.855650830 +-14.90% 0.887557919 +-14.90%
dd write 200 MiB 1.688322953 +-12.72% 1.667682724 +-13.33%
dd write 1000 MiB 8.418601605 +-14.30% 8.673532299 +-15.00%
read fault cycles 266.0(+-0.000% 10 samples) 266.0(+-0.000% 10 samples)
write fault cycles 2051.7(+-1.349% 10 samples) 2049.6(+-1.686% 10 samples)
* CONFIG_MEMCG=y non-root_memcg:
baseline patched
kbuild 1m26.120000(+-0.273% 3 samples) 1m25.763333(+-0.127% 3 samples)
dd write 100 MiB 0.861723964 +-15.25% 0.818129350 +-14.82%
dd write 200 MiB 1.669887569 +-13.30% 1.698645885 +-13.27%
dd write 1000 MiB 8.383191730 +-14.65% 8.351742280 +-14.52%
read fault cycles 265.7(+-0.172% 10 samples) 267.0(+-0.000% 10 samples)
write fault cycles 2070.6(+-1.512% 10 samples) 2084.4(+-2.148% 10 samples)
As expected anon page faults are not affected by this patch.
tj: Updated to apply on top of the recent cancel_dirty_page() changes.
Signed-off-by: Sha Zhengju <handai.szj@gmail.com>
Signed-off-by: Greg Thelen <gthelen@google.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
2015-05-22 21:13:16 +00:00
|
|
|
|
2021-12-23 21:50:23 +00:00
|
|
|
BUG_ON(!folio_test_locked(folio));
|
|
|
|
BUG_ON(mapping != folio_mapping(folio));
|
2006-01-08 09:00:48 +00:00
|
|
|
|
2021-12-23 21:50:23 +00:00
|
|
|
if (!folio_test_swapcache(folio))
|
vfs: keep inodes with page cache off the inode shrinker LRU
Historically (pre-2.5), the inode shrinker used to reclaim only empty
inodes and skip over those that still contained page cache. This caused
problems on highmem hosts: struct inode could put fill lowmem zones
before the cache was getting reclaimed in the highmem zones.
To address this, the inode shrinker started to strip page cache to
facilitate reclaiming lowmem. However, this comes with its own set of
problems: the shrinkers may drop actively used page cache just because
the inodes are not currently open or dirty - think working with a large
git tree. It further doesn't respect cgroup memory protection settings
and can cause priority inversions between containers.
Nowadays, the page cache also holds non-resident info for evicted cache
pages in order to detect refaults. We've come to rely heavily on this
data inside reclaim for protecting the cache workingset and driving swap
behavior. We also use it to quantify and report workload health through
psi. The latter in turn is used for fleet health monitoring, as well as
driving automated memory sizing of workloads and containers, proactive
reclaim and memory offloading schemes.
The consequences of dropping page cache prematurely is that we're seeing
subtle and not-so-subtle failures in all of the above-mentioned
scenarios, with the workload generally entering unexpected thrashing
states while losing the ability to reliably detect it.
To fix this on non-highmem systems at least, going back to rotating
inodes on the LRU isn't feasible. We've tried (commit a76cf1a474d7
("mm: don't reclaim inodes with many attached pages")) and failed
(commit 69056ee6a8a3 ("Revert "mm: don't reclaim inodes with many
attached pages"")).
The issue is mostly that shrinker pools attract pressure based on their
size, and when objects get skipped the shrinkers remember this as
deferred reclaim work. This accumulates excessive pressure on the
remaining inodes, and we can quickly eat into heavily used ones, or
dirty ones that require IO to reclaim, when there potentially is plenty
of cold, clean cache around still.
Instead, this patch keeps populated inodes off the inode LRU in the
first place - just like an open file or dirty state would. An otherwise
clean and unused inode then gets queued when the last cache entry
disappears. This solves the problem without reintroducing the reclaim
issues, and generally is a bit more scalable than having to wade through
potentially hundreds of thousands of busy inodes.
Locking is a bit tricky because the locks protecting the inode state
(i_lock) and the inode LRU (lru_list.lock) don't nest inside the
irq-safe page cache lock (i_pages.xa_lock). Page cache deletions are
serialized through i_lock, taken before the i_pages lock, to make sure
depopulated inodes are queued reliably. Additions may race with
deletions, but we'll check again in the shrinker. If additions race
with the shrinker itself, we're protected by the i_lock: if find_inode()
or iput() win, the shrinker will bail on the elevated i_count or
I_REFERENCED; if the shrinker wins and goes ahead with the inode, it
will set I_FREEING and inhibit further igets(), which will cause the
other side to create a new instance of the inode instead.
Link: https://lkml.kernel.org/r/20210614211904.14420-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-09 02:31:24 +00:00
|
|
|
spin_lock(&mapping->host->i_lock);
|
2021-09-02 21:53:18 +00:00
|
|
|
xa_lock_irq(&mapping->i_pages);
|
2006-01-08 09:00:48 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* The non racy check for a busy folio.
|
2006-09-27 08:50:02 +00:00
|
|
|
*
|
|
|
|
* Must be careful with the order of the tests. When someone has
|
2022-09-02 19:45:57 +00:00
|
|
|
* a ref to the folio, it may be possible that they dirty it then
|
|
|
|
* drop the reference. So if the dirty flag is tested before the
|
|
|
|
* refcount here, then the following race may occur:
|
2006-09-27 08:50:02 +00:00
|
|
|
*
|
|
|
|
* get_user_pages(&page);
|
|
|
|
* [user mapping goes away]
|
|
|
|
* write_to(page);
|
2022-09-02 19:45:57 +00:00
|
|
|
* !folio_test_dirty(folio) [good]
|
|
|
|
* folio_set_dirty(folio);
|
|
|
|
* folio_put(folio);
|
|
|
|
* !refcount(folio) [good, discard it]
|
2006-09-27 08:50:02 +00:00
|
|
|
*
|
|
|
|
* [oops, our write_to data is lost]
|
|
|
|
*
|
|
|
|
* Reversing the order of the tests ensures such a situation cannot
|
2022-09-02 19:45:57 +00:00
|
|
|
* escape unnoticed. The smp_rmb is needed to ensure the folio->flags
|
|
|
|
* load is not satisfied before that of folio->_refcount.
|
2006-09-27 08:50:02 +00:00
|
|
|
*
|
2022-09-02 19:45:57 +00:00
|
|
|
* Note that if the dirty flag is always set via folio_mark_dirty,
|
2018-04-10 23:36:56 +00:00
|
|
|
* and thus under the i_pages lock, then this ordering is not required.
|
2006-01-08 09:00:48 +00:00
|
|
|
*/
|
2021-12-23 21:50:23 +00:00
|
|
|
refcount = 1 + folio_nr_pages(folio);
|
|
|
|
if (!folio_ref_freeze(folio, refcount))
|
2006-01-08 09:00:48 +00:00
|
|
|
goto cannot_free;
|
2022-09-02 19:45:57 +00:00
|
|
|
/* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
|
2021-12-23 21:50:23 +00:00
|
|
|
if (unlikely(folio_test_dirty(folio))) {
|
|
|
|
folio_ref_unfreeze(folio, refcount);
|
2006-01-08 09:00:48 +00:00
|
|
|
goto cannot_free;
|
2008-07-26 02:45:30 +00:00
|
|
|
}
|
2006-01-08 09:00:48 +00:00
|
|
|
|
2021-12-23 21:50:23 +00:00
|
|
|
if (folio_test_swapcache(folio)) {
|
|
|
|
swp_entry_t swap = folio_swap_entry(folio);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
2020-08-12 01:30:50 +00:00
|
|
|
if (reclaimed && !mapping_exiting(mapping))
|
2021-12-23 21:39:05 +00:00
|
|
|
shadow = workingset_eviction(folio, target_memcg);
|
2022-06-17 17:50:20 +00:00
|
|
|
__delete_from_swap_cache(folio, swap, shadow);
|
2022-12-05 01:01:03 +00:00
|
|
|
mem_cgroup_swapout(folio, swap);
|
2021-09-02 21:53:18 +00:00
|
|
|
xa_unlock_irq(&mapping->i_pages);
|
2022-09-02 19:46:09 +00:00
|
|
|
put_swap_folio(folio, swap);
|
2008-07-26 02:45:30 +00:00
|
|
|
} else {
|
2022-05-01 11:35:31 +00:00
|
|
|
void (*free_folio)(struct folio *);
|
2010-12-01 18:35:19 +00:00
|
|
|
|
2022-05-01 11:35:31 +00:00
|
|
|
free_folio = mapping->a_ops->free_folio;
|
2014-04-03 21:47:51 +00:00
|
|
|
/*
|
|
|
|
* Remember a shadow entry for reclaimed file cache in
|
|
|
|
* order to detect refaults, thus thrashing, later on.
|
|
|
|
*
|
|
|
|
* But don't store shadows in an address space that is
|
2020-08-07 06:26:29 +00:00
|
|
|
* already exiting. This is not just an optimization,
|
2014-04-03 21:47:51 +00:00
|
|
|
* inode reclaim needs to empty out the radix tree or
|
|
|
|
* the nodes are lost. Don't plant shadows behind its
|
|
|
|
* back.
|
2016-01-22 23:10:40 +00:00
|
|
|
*
|
|
|
|
* We also don't store shadows for DAX mappings because the
|
2022-09-02 19:45:57 +00:00
|
|
|
* only page cache folios found in these are zero pages
|
2016-01-22 23:10:40 +00:00
|
|
|
* covering holes, and because we don't want to mix DAX
|
|
|
|
* exceptional entries and shadow exceptional entries in the
|
2018-04-10 23:36:56 +00:00
|
|
|
* same address_space.
|
2014-04-03 21:47:51 +00:00
|
|
|
*/
|
2021-12-23 21:50:23 +00:00
|
|
|
if (reclaimed && folio_is_file_lru(folio) &&
|
2016-01-22 23:10:40 +00:00
|
|
|
!mapping_exiting(mapping) && !dax_mapping(mapping))
|
2021-12-23 21:39:05 +00:00
|
|
|
shadow = workingset_eviction(folio, target_memcg);
|
|
|
|
__filemap_remove_folio(folio, shadow);
|
2021-09-02 21:53:18 +00:00
|
|
|
xa_unlock_irq(&mapping->i_pages);
|
vfs: keep inodes with page cache off the inode shrinker LRU
Historically (pre-2.5), the inode shrinker used to reclaim only empty
inodes and skip over those that still contained page cache. This caused
problems on highmem hosts: struct inode could put fill lowmem zones
before the cache was getting reclaimed in the highmem zones.
To address this, the inode shrinker started to strip page cache to
facilitate reclaiming lowmem. However, this comes with its own set of
problems: the shrinkers may drop actively used page cache just because
the inodes are not currently open or dirty - think working with a large
git tree. It further doesn't respect cgroup memory protection settings
and can cause priority inversions between containers.
Nowadays, the page cache also holds non-resident info for evicted cache
pages in order to detect refaults. We've come to rely heavily on this
data inside reclaim for protecting the cache workingset and driving swap
behavior. We also use it to quantify and report workload health through
psi. The latter in turn is used for fleet health monitoring, as well as
driving automated memory sizing of workloads and containers, proactive
reclaim and memory offloading schemes.
The consequences of dropping page cache prematurely is that we're seeing
subtle and not-so-subtle failures in all of the above-mentioned
scenarios, with the workload generally entering unexpected thrashing
states while losing the ability to reliably detect it.
To fix this on non-highmem systems at least, going back to rotating
inodes on the LRU isn't feasible. We've tried (commit a76cf1a474d7
("mm: don't reclaim inodes with many attached pages")) and failed
(commit 69056ee6a8a3 ("Revert "mm: don't reclaim inodes with many
attached pages"")).
The issue is mostly that shrinker pools attract pressure based on their
size, and when objects get skipped the shrinkers remember this as
deferred reclaim work. This accumulates excessive pressure on the
remaining inodes, and we can quickly eat into heavily used ones, or
dirty ones that require IO to reclaim, when there potentially is plenty
of cold, clean cache around still.
Instead, this patch keeps populated inodes off the inode LRU in the
first place - just like an open file or dirty state would. An otherwise
clean and unused inode then gets queued when the last cache entry
disappears. This solves the problem without reintroducing the reclaim
issues, and generally is a bit more scalable than having to wade through
potentially hundreds of thousands of busy inodes.
Locking is a bit tricky because the locks protecting the inode state
(i_lock) and the inode LRU (lru_list.lock) don't nest inside the
irq-safe page cache lock (i_pages.xa_lock). Page cache deletions are
serialized through i_lock, taken before the i_pages lock, to make sure
depopulated inodes are queued reliably. Additions may race with
deletions, but we'll check again in the shrinker. If additions race
with the shrinker itself, we're protected by the i_lock: if find_inode()
or iput() win, the shrinker will bail on the elevated i_count or
I_REFERENCED; if the shrinker wins and goes ahead with the inode, it
will set I_FREEING and inhibit further igets(), which will cause the
other side to create a new instance of the inode instead.
Link: https://lkml.kernel.org/r/20210614211904.14420-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-09 02:31:24 +00:00
|
|
|
if (mapping_shrinkable(mapping))
|
|
|
|
inode_add_lru(mapping->host);
|
|
|
|
spin_unlock(&mapping->host->i_lock);
|
2010-12-01 18:35:19 +00:00
|
|
|
|
2022-05-01 11:35:31 +00:00
|
|
|
if (free_folio)
|
|
|
|
free_folio(folio);
|
2006-01-08 09:00:48 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
cannot_free:
|
2021-09-02 21:53:18 +00:00
|
|
|
xa_unlock_irq(&mapping->i_pages);
|
2021-12-23 21:50:23 +00:00
|
|
|
if (!folio_test_swapcache(folio))
|
vfs: keep inodes with page cache off the inode shrinker LRU
Historically (pre-2.5), the inode shrinker used to reclaim only empty
inodes and skip over those that still contained page cache. This caused
problems on highmem hosts: struct inode could put fill lowmem zones
before the cache was getting reclaimed in the highmem zones.
To address this, the inode shrinker started to strip page cache to
facilitate reclaiming lowmem. However, this comes with its own set of
problems: the shrinkers may drop actively used page cache just because
the inodes are not currently open or dirty - think working with a large
git tree. It further doesn't respect cgroup memory protection settings
and can cause priority inversions between containers.
Nowadays, the page cache also holds non-resident info for evicted cache
pages in order to detect refaults. We've come to rely heavily on this
data inside reclaim for protecting the cache workingset and driving swap
behavior. We also use it to quantify and report workload health through
psi. The latter in turn is used for fleet health monitoring, as well as
driving automated memory sizing of workloads and containers, proactive
reclaim and memory offloading schemes.
The consequences of dropping page cache prematurely is that we're seeing
subtle and not-so-subtle failures in all of the above-mentioned
scenarios, with the workload generally entering unexpected thrashing
states while losing the ability to reliably detect it.
To fix this on non-highmem systems at least, going back to rotating
inodes on the LRU isn't feasible. We've tried (commit a76cf1a474d7
("mm: don't reclaim inodes with many attached pages")) and failed
(commit 69056ee6a8a3 ("Revert "mm: don't reclaim inodes with many
attached pages"")).
The issue is mostly that shrinker pools attract pressure based on their
size, and when objects get skipped the shrinkers remember this as
deferred reclaim work. This accumulates excessive pressure on the
remaining inodes, and we can quickly eat into heavily used ones, or
dirty ones that require IO to reclaim, when there potentially is plenty
of cold, clean cache around still.
Instead, this patch keeps populated inodes off the inode LRU in the
first place - just like an open file or dirty state would. An otherwise
clean and unused inode then gets queued when the last cache entry
disappears. This solves the problem without reintroducing the reclaim
issues, and generally is a bit more scalable than having to wade through
potentially hundreds of thousands of busy inodes.
Locking is a bit tricky because the locks protecting the inode state
(i_lock) and the inode LRU (lru_list.lock) don't nest inside the
irq-safe page cache lock (i_pages.xa_lock). Page cache deletions are
serialized through i_lock, taken before the i_pages lock, to make sure
depopulated inodes are queued reliably. Additions may race with
deletions, but we'll check again in the shrinker. If additions race
with the shrinker itself, we're protected by the i_lock: if find_inode()
or iput() win, the shrinker will bail on the elevated i_count or
I_REFERENCED; if the shrinker wins and goes ahead with the inode, it
will set I_FREEING and inhibit further igets(), which will cause the
other side to create a new instance of the inode instead.
Link: https://lkml.kernel.org/r/20210614211904.14420-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Roman Gushchin <guro@fb.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-09 02:31:24 +00:00
|
|
|
spin_unlock(&mapping->host->i_lock);
|
2006-01-08 09:00:48 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2022-02-13 03:48:55 +00:00
|
|
|
/**
|
|
|
|
* remove_mapping() - Attempt to remove a folio from its mapping.
|
|
|
|
* @mapping: The address space.
|
|
|
|
* @folio: The folio to remove.
|
|
|
|
*
|
|
|
|
* If the folio is dirty, under writeback or if someone else has a ref
|
|
|
|
* on it, removal will fail.
|
|
|
|
* Return: The number of pages removed from the mapping. 0 if the folio
|
|
|
|
* could not be removed.
|
|
|
|
* Context: The caller should have a single refcount on the folio and
|
|
|
|
* hold its lock.
|
2008-07-26 02:45:30 +00:00
|
|
|
*/
|
2022-02-13 03:48:55 +00:00
|
|
|
long remove_mapping(struct address_space *mapping, struct folio *folio)
|
2008-07-26 02:45:30 +00:00
|
|
|
{
|
2021-12-23 21:50:23 +00:00
|
|
|
if (__remove_mapping(mapping, folio, false, NULL)) {
|
2008-07-26 02:45:30 +00:00
|
|
|
/*
|
2022-02-13 03:48:55 +00:00
|
|
|
* Unfreezing the refcount with 1 effectively
|
2008-07-26 02:45:30 +00:00
|
|
|
* drops the pagecache ref for us without requiring another
|
|
|
|
* atomic operation.
|
|
|
|
*/
|
2021-12-23 21:50:23 +00:00
|
|
|
folio_ref_unfreeze(folio, 1);
|
2022-02-13 03:48:55 +00:00
|
|
|
return folio_nr_pages(folio);
|
2008-07-26 02:45:30 +00:00
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
/**
|
2022-01-21 13:41:46 +00:00
|
|
|
* folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
|
|
|
|
* @folio: Folio to be returned to an LRU list.
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
*
|
2022-01-21 13:41:46 +00:00
|
|
|
* Add previously isolated @folio to appropriate LRU list.
|
|
|
|
* The folio may still be unevictable for other reasons.
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
*
|
2022-01-21 13:41:46 +00:00
|
|
|
* Context: lru_lock must not be held, interrupts must be enabled.
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
*/
|
2022-01-21 13:41:46 +00:00
|
|
|
void folio_putback_lru(struct folio *folio)
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
{
|
2022-01-21 13:41:46 +00:00
|
|
|
folio_add_lru(folio);
|
|
|
|
folio_put(folio); /* drop ref from isolate */
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
}
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
enum folio_references {
|
|
|
|
FOLIOREF_RECLAIM,
|
|
|
|
FOLIOREF_RECLAIM_CLEAN,
|
|
|
|
FOLIOREF_KEEP,
|
|
|
|
FOLIOREF_ACTIVATE,
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
};
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
static enum folio_references folio_check_references(struct folio *folio,
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
struct scan_control *sc)
|
|
|
|
{
|
2022-02-15 18:44:40 +00:00
|
|
|
int referenced_ptes, referenced_folio;
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
unsigned long vm_flags;
|
|
|
|
|
2022-01-21 16:27:31 +00:00
|
|
|
referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
|
|
|
|
&vm_flags);
|
2022-02-15 18:44:40 +00:00
|
|
|
referenced_folio = folio_test_clear_referenced(folio);
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
|
|
|
|
/*
|
2022-02-15 18:44:40 +00:00
|
|
|
* The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
|
|
|
|
* Let the folio, now marked Mlocked, be moved to the unevictable list.
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
*/
|
|
|
|
if (vm_flags & VM_LOCKED)
|
2022-09-02 19:45:57 +00:00
|
|
|
return FOLIOREF_ACTIVATE;
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
|
mm: don't be stuck to rmap lock on reclaim path
The rmap locks(i_mmap_rwsem and anon_vma->root->rwsem) could be contended
under memory pressure if processes keep working on their vmas(e.g., fork,
mmap, munmap). It makes reclaim path stuck. In our real workload traces,
we see kswapd is waiting the lock for 300ms+(worst case, a sec) and it
makes other processes entering direct reclaim, which were also stuck on
the lock.
This patch makes lru aging path try_lock mode like shink_page_list so the
reclaim context will keep working with next lru pages without being stuck.
if it found the rmap lock contended, it rotates the page back to head of
lru in both active/inactive lrus to make them consistent behavior, which
is basic starting point rather than adding more heristic.
Since this patch introduces a new "contended" field as out-param along
with try_lock in-param in rmap_walk_control, it's not immutable any longer
if the try_lock is set so remove const keywords on rmap related functions.
Since rmap walking is already expensive operation, I doubt the const
would help sizable benefit( And we didn't have it until 5.17).
In a heavy app workload in Android, trace shows following statistics. It
almost removes rmap lock contention from reclaim path.
Martin Liu reported:
Before:
max_dur(ms) min_dur(ms) max-min(dur)ms avg_dur(ms) sum_dur(ms) count blocked_function
1632 0 1631 151.542173 31672 209 page_lock_anon_vma_read
601 0 601 145.544681 28817 198 rmap_walk_file
After:
max_dur(ms) min_dur(ms) max-min(dur)ms avg_dur(ms) sum_dur(ms) count blocked_function
NaN NaN NaN NaN NaN 0.0 NaN
0 0 0 0.127645 1 12 rmap_walk_file
[minchan@kernel.org: add comment, per Matthew]
Link: https://lkml.kernel.org/r/YnNqeB5tUf6LZ57b@google.com
Link: https://lkml.kernel.org/r/20220510215423.164547-1-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: John Dias <joaodias@google.com>
Cc: Tim Murray <timmurray@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Martin Liu <liumartin@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-19 21:08:54 +00:00
|
|
|
/* rmap lock contention: rotate */
|
|
|
|
if (referenced_ptes == -1)
|
2022-09-02 19:45:57 +00:00
|
|
|
return FOLIOREF_KEEP;
|
mm: don't be stuck to rmap lock on reclaim path
The rmap locks(i_mmap_rwsem and anon_vma->root->rwsem) could be contended
under memory pressure if processes keep working on their vmas(e.g., fork,
mmap, munmap). It makes reclaim path stuck. In our real workload traces,
we see kswapd is waiting the lock for 300ms+(worst case, a sec) and it
makes other processes entering direct reclaim, which were also stuck on
the lock.
This patch makes lru aging path try_lock mode like shink_page_list so the
reclaim context will keep working with next lru pages without being stuck.
if it found the rmap lock contended, it rotates the page back to head of
lru in both active/inactive lrus to make them consistent behavior, which
is basic starting point rather than adding more heristic.
Since this patch introduces a new "contended" field as out-param along
with try_lock in-param in rmap_walk_control, it's not immutable any longer
if the try_lock is set so remove const keywords on rmap related functions.
Since rmap walking is already expensive operation, I doubt the const
would help sizable benefit( And we didn't have it until 5.17).
In a heavy app workload in Android, trace shows following statistics. It
almost removes rmap lock contention from reclaim path.
Martin Liu reported:
Before:
max_dur(ms) min_dur(ms) max-min(dur)ms avg_dur(ms) sum_dur(ms) count blocked_function
1632 0 1631 151.542173 31672 209 page_lock_anon_vma_read
601 0 601 145.544681 28817 198 rmap_walk_file
After:
max_dur(ms) min_dur(ms) max-min(dur)ms avg_dur(ms) sum_dur(ms) count blocked_function
NaN NaN NaN NaN NaN 0.0 NaN
0 0 0 0.127645 1 12 rmap_walk_file
[minchan@kernel.org: add comment, per Matthew]
Link: https://lkml.kernel.org/r/YnNqeB5tUf6LZ57b@google.com
Link: https://lkml.kernel.org/r/20220510215423.164547-1-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: John Dias <joaodias@google.com>
Cc: Tim Murray <timmurray@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Martin Liu <liumartin@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-19 21:08:54 +00:00
|
|
|
|
2010-03-05 21:42:22 +00:00
|
|
|
if (referenced_ptes) {
|
|
|
|
/*
|
2022-02-15 18:44:40 +00:00
|
|
|
* All mapped folios start out with page table
|
2010-03-05 21:42:22 +00:00
|
|
|
* references from the instantiating fault, so we need
|
2022-03-23 00:03:12 +00:00
|
|
|
* to look twice if a mapped file/anon folio is used more
|
2010-03-05 21:42:22 +00:00
|
|
|
* than once.
|
|
|
|
*
|
|
|
|
* Mark it and spare it for another trip around the
|
|
|
|
* inactive list. Another page table reference will
|
|
|
|
* lead to its activation.
|
|
|
|
*
|
2022-02-15 18:44:40 +00:00
|
|
|
* Note: the mark is set for activated folios as well
|
|
|
|
* so that recently deactivated but used folios are
|
2010-03-05 21:42:22 +00:00
|
|
|
* quickly recovered.
|
|
|
|
*/
|
2022-02-15 18:44:40 +00:00
|
|
|
folio_set_referenced(folio);
|
2010-03-05 21:42:22 +00:00
|
|
|
|
2022-02-15 18:44:40 +00:00
|
|
|
if (referenced_folio || referenced_ptes > 1)
|
2022-09-02 19:45:57 +00:00
|
|
|
return FOLIOREF_ACTIVATE;
|
2010-03-05 21:42:22 +00:00
|
|
|
|
2012-01-10 23:07:03 +00:00
|
|
|
/*
|
2022-02-15 18:44:40 +00:00
|
|
|
* Activate file-backed executable folios after first usage.
|
2012-01-10 23:07:03 +00:00
|
|
|
*/
|
2022-05-13 03:23:00 +00:00
|
|
|
if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
|
2022-09-02 19:45:57 +00:00
|
|
|
return FOLIOREF_ACTIVATE;
|
2012-01-10 23:07:03 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
return FOLIOREF_KEEP;
|
2010-03-05 21:42:22 +00:00
|
|
|
}
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
|
2022-02-15 18:44:40 +00:00
|
|
|
/* Reclaim if clean, defer dirty folios to writeback */
|
2022-05-13 03:23:00 +00:00
|
|
|
if (referenced_folio && folio_is_file_lru(folio))
|
2022-09-02 19:45:57 +00:00
|
|
|
return FOLIOREF_RECLAIM_CLEAN;
|
2010-03-05 21:42:22 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
return FOLIOREF_RECLAIM;
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
}
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
/* Check if a folio is dirty or under writeback */
|
2022-01-17 19:35:22 +00:00
|
|
|
static void folio_check_dirty_writeback(struct folio *folio,
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
bool *dirty, bool *writeback)
|
|
|
|
{
|
2013-07-03 22:02:05 +00:00
|
|
|
struct address_space *mapping;
|
|
|
|
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* Anonymous folios are not handled by flushers and must be written
|
2022-05-13 03:22:59 +00:00
|
|
|
* from reclaim context. Do not stall reclaim based on them.
|
2022-09-02 19:45:57 +00:00
|
|
|
* MADV_FREE anonymous folios are put into inactive file list too.
|
2022-05-13 03:22:59 +00:00
|
|
|
* They could be mistakenly treated as file lru. So further anon
|
|
|
|
* test is needed.
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
*/
|
2022-01-17 19:35:22 +00:00
|
|
|
if (!folio_is_file_lru(folio) ||
|
|
|
|
(folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
*dirty = false;
|
|
|
|
*writeback = false;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2022-01-17 19:35:22 +00:00
|
|
|
/* By default assume that the folio flags are accurate */
|
|
|
|
*dirty = folio_test_dirty(folio);
|
|
|
|
*writeback = folio_test_writeback(folio);
|
2013-07-03 22:02:05 +00:00
|
|
|
|
|
|
|
/* Verify dirty/writeback state if the filesystem supports it */
|
2022-01-17 19:35:22 +00:00
|
|
|
if (!folio_test_private(folio))
|
2013-07-03 22:02:05 +00:00
|
|
|
return;
|
|
|
|
|
2022-01-17 19:35:22 +00:00
|
|
|
mapping = folio_mapping(folio);
|
2013-07-03 22:02:05 +00:00
|
|
|
if (mapping && mapping->a_ops->is_dirty_writeback)
|
2022-01-17 19:35:22 +00:00
|
|
|
mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
}
|
|
|
|
|
2022-08-18 13:10:40 +00:00
|
|
|
static struct page *alloc_demote_page(struct page *page, unsigned long private)
|
2021-09-02 21:59:16 +00:00
|
|
|
{
|
2022-08-18 13:10:40 +00:00
|
|
|
struct page *target_page;
|
|
|
|
nodemask_t *allowed_mask;
|
|
|
|
struct migration_target_control *mtc;
|
|
|
|
|
|
|
|
mtc = (struct migration_target_control *)private;
|
|
|
|
|
|
|
|
allowed_mask = mtc->nmask;
|
|
|
|
/*
|
|
|
|
* make sure we allocate from the target node first also trying to
|
|
|
|
* demote or reclaim pages from the target node via kswapd if we are
|
|
|
|
* low on free memory on target node. If we don't do this and if
|
|
|
|
* we have free memory on the slower(lower) memtier, we would start
|
|
|
|
* allocating pages from slower(lower) memory tiers without even forcing
|
|
|
|
* a demotion of cold pages from the target memtier. This can result
|
|
|
|
* in the kernel placing hot pages in slower(lower) memory tiers.
|
|
|
|
*/
|
|
|
|
mtc->nmask = NULL;
|
|
|
|
mtc->gfp_mask |= __GFP_THISNODE;
|
|
|
|
target_page = alloc_migration_target(page, (unsigned long)mtc);
|
|
|
|
if (target_page)
|
|
|
|
return target_page;
|
2021-09-02 21:59:16 +00:00
|
|
|
|
2022-08-18 13:10:40 +00:00
|
|
|
mtc->gfp_mask &= ~__GFP_THISNODE;
|
|
|
|
mtc->nmask = allowed_mask;
|
|
|
|
|
|
|
|
return alloc_migration_target(page, (unsigned long)mtc);
|
2021-09-02 21:59:16 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* Take folios on @demote_folios and attempt to demote them to another node.
|
|
|
|
* Folios which are not demoted are left on @demote_folios.
|
2021-09-02 21:59:16 +00:00
|
|
|
*/
|
2022-09-02 19:45:57 +00:00
|
|
|
static unsigned int demote_folio_list(struct list_head *demote_folios,
|
2021-09-02 21:59:16 +00:00
|
|
|
struct pglist_data *pgdat)
|
|
|
|
{
|
|
|
|
int target_nid = next_demotion_node(pgdat->node_id);
|
|
|
|
unsigned int nr_succeeded;
|
2022-08-18 13:10:40 +00:00
|
|
|
nodemask_t allowed_mask;
|
|
|
|
|
|
|
|
struct migration_target_control mtc = {
|
|
|
|
/*
|
|
|
|
* Allocate from 'node', or fail quickly and quietly.
|
|
|
|
* When this happens, 'page' will likely just be discarded
|
|
|
|
* instead of migrated.
|
|
|
|
*/
|
|
|
|
.gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
|
|
|
|
__GFP_NOMEMALLOC | GFP_NOWAIT,
|
|
|
|
.nid = target_nid,
|
|
|
|
.nmask = &allowed_mask
|
|
|
|
};
|
2021-09-02 21:59:16 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
if (list_empty(demote_folios))
|
2021-09-02 21:59:16 +00:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (target_nid == NUMA_NO_NODE)
|
|
|
|
return 0;
|
|
|
|
|
2022-08-18 13:10:40 +00:00
|
|
|
node_get_allowed_targets(pgdat, &allowed_mask);
|
|
|
|
|
2021-09-02 21:59:16 +00:00
|
|
|
/* Demotion ignores all cpuset and mempolicy settings */
|
2022-09-02 19:45:57 +00:00
|
|
|
migrate_pages(demote_folios, alloc_demote_page, NULL,
|
2022-08-18 13:10:40 +00:00
|
|
|
(unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
|
|
|
|
&nr_succeeded);
|
2021-09-02 21:59:16 +00:00
|
|
|
|
2022-10-26 18:01:33 +00:00
|
|
|
__count_vm_events(PGDEMOTE_KSWAPD + reclaimer_offset(), nr_succeeded);
|
2021-09-02 21:59:19 +00:00
|
|
|
|
2021-09-02 21:59:16 +00:00
|
|
|
return nr_succeeded;
|
|
|
|
}
|
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
|
2022-05-10 01:20:48 +00:00
|
|
|
{
|
|
|
|
if (gfp_mask & __GFP_FS)
|
|
|
|
return true;
|
2022-05-13 03:23:03 +00:00
|
|
|
if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
|
2022-05-10 01:20:48 +00:00
|
|
|
return false;
|
|
|
|
/*
|
|
|
|
* We can "enter_fs" for swap-cache with only __GFP_IO
|
|
|
|
* providing this isn't SWP_FS_OPS.
|
|
|
|
* ->flags can be updated non-atomicially (scan_swap_map_slots),
|
|
|
|
* but that will never affect SWP_FS_OPS, so the data_race
|
|
|
|
* is safe.
|
|
|
|
*/
|
2022-06-17 17:50:18 +00:00
|
|
|
return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
|
2022-05-10 01:20:48 +00:00
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* shrink_folio_list() returns the number of reclaimed pages
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2022-09-02 19:45:57 +00:00
|
|
|
static unsigned int shrink_folio_list(struct list_head *folio_list,
|
|
|
|
struct pglist_data *pgdat, struct scan_control *sc,
|
|
|
|
struct reclaim_stat *stat, bool ignore_references)
|
|
|
|
{
|
|
|
|
LIST_HEAD(ret_folios);
|
|
|
|
LIST_HEAD(free_folios);
|
|
|
|
LIST_HEAD(demote_folios);
|
2020-06-03 23:01:18 +00:00
|
|
|
unsigned int nr_reclaimed = 0;
|
|
|
|
unsigned int pgactivate = 0;
|
2021-09-02 21:59:16 +00:00
|
|
|
bool do_demote_pass;
|
2022-05-10 01:20:49 +00:00
|
|
|
struct swap_iocb *plug = NULL;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2019-03-05 23:48:15 +00:00
|
|
|
memset(stat, 0, sizeof(*stat));
|
2005-04-16 22:20:36 +00:00
|
|
|
cond_resched();
|
2021-09-02 21:59:16 +00:00
|
|
|
do_demote_pass = can_demote(pgdat->node_id, sc);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2021-09-02 21:59:16 +00:00
|
|
|
retry:
|
2022-09-02 19:45:57 +00:00
|
|
|
while (!list_empty(folio_list)) {
|
2005-04-16 22:20:36 +00:00
|
|
|
struct address_space *mapping;
|
2021-12-23 21:50:23 +00:00
|
|
|
struct folio *folio;
|
2022-09-02 19:45:57 +00:00
|
|
|
enum folio_references references = FOLIOREF_RECLAIM;
|
2022-05-10 01:20:48 +00:00
|
|
|
bool dirty, writeback;
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
unsigned int nr_pages;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
cond_resched();
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
folio = lru_to_folio(folio_list);
|
2021-12-23 21:50:23 +00:00
|
|
|
list_del(&folio->lru);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
if (!folio_trylock(folio))
|
2005-04-16 22:20:36 +00:00
|
|
|
goto keep;
|
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
nr_pages = folio_nr_pages(folio);
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
/* Account the number of base pages */
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
sc->nr_scanned += nr_pages;
|
2006-02-12 01:55:53 +00:00
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
if (unlikely(!folio_evictable(folio)))
|
2017-05-03 21:54:13 +00:00
|
|
|
goto activate_locked;
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
|
2022-05-13 03:23:01 +00:00
|
|
|
if (!sc->may_unmap && folio_mapped(folio))
|
2006-02-12 01:55:53 +00:00
|
|
|
goto keep_locked;
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
/* folio_update_gen() tried to promote this page? */
|
|
|
|
if (lru_gen_enabled() && !ignore_references &&
|
|
|
|
folio_mapped(folio) && folio_test_referenced(folio))
|
|
|
|
goto keep_locked;
|
|
|
|
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
/*
|
2018-04-10 23:27:51 +00:00
|
|
|
* The number of dirty pages determines if a node is marked
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
* reclaim_congested. kswapd will stall and start writing
|
2022-05-13 03:23:03 +00:00
|
|
|
* folios if the tail of the LRU is all dirty unqueued folios.
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
*/
|
2022-01-17 19:35:22 +00:00
|
|
|
folio_check_dirty_writeback(folio, &dirty, &writeback);
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
if (dirty || writeback)
|
2022-01-17 19:46:55 +00:00
|
|
|
stat->nr_dirty += nr_pages;
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
|
|
|
|
if (dirty && !writeback)
|
2022-01-17 19:46:55 +00:00
|
|
|
stat->nr_unqueued_dirty += nr_pages;
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
|
2013-07-03 22:02:03 +00:00
|
|
|
/*
|
2022-05-13 03:23:03 +00:00
|
|
|
* Treat this folio as congested if folios are cycling
|
|
|
|
* through the LRU so quickly that the folios marked
|
|
|
|
* for immediate reclaim are making it to the end of
|
|
|
|
* the LRU a second time.
|
2013-07-03 22:02:03 +00:00
|
|
|
*/
|
2022-05-13 03:23:03 +00:00
|
|
|
if (writeback && folio_test_reclaim(folio))
|
2022-01-17 19:46:55 +00:00
|
|
|
stat->nr_congested += nr_pages;
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
|
2013-07-03 22:01:51 +00:00
|
|
|
/*
|
2022-05-13 03:23:02 +00:00
|
|
|
* If a folio at the tail of the LRU is under writeback, there
|
2013-07-03 22:01:51 +00:00
|
|
|
* are three cases to consider.
|
|
|
|
*
|
2022-05-13 03:23:03 +00:00
|
|
|
* 1) If reclaim is encountering an excessive number
|
|
|
|
* of folios under writeback and this folio has both
|
|
|
|
* the writeback and reclaim flags set, then it
|
2022-05-13 03:23:02 +00:00
|
|
|
* indicates that folios are being queued for I/O but
|
|
|
|
* are being recycled through the LRU before the I/O
|
|
|
|
* can complete. Waiting on the folio itself risks an
|
|
|
|
* indefinite stall if it is impossible to writeback
|
|
|
|
* the folio due to I/O error or disconnected storage
|
|
|
|
* so instead note that the LRU is being scanned too
|
|
|
|
* quickly and the caller can stall after the folio
|
|
|
|
* list has been processed.
|
2013-07-03 22:01:51 +00:00
|
|
|
*
|
2022-05-13 03:23:02 +00:00
|
|
|
* 2) Global or new memcg reclaim encounters a folio that is
|
2015-08-04 21:36:58 +00:00
|
|
|
* not marked for immediate reclaim, or the caller does not
|
|
|
|
* have __GFP_FS (or __GFP_IO if it's simply going to swap,
|
2022-05-13 03:23:02 +00:00
|
|
|
* not to fs). In this case mark the folio for immediate
|
2015-05-22 22:23:36 +00:00
|
|
|
* reclaim and continue scanning.
|
2013-07-03 22:01:51 +00:00
|
|
|
*
|
2022-05-10 01:20:48 +00:00
|
|
|
* Require may_enter_fs() because we would wait on fs, which
|
2022-05-13 03:23:02 +00:00
|
|
|
* may not have submitted I/O yet. And the loop driver might
|
|
|
|
* enter reclaim, and deadlock if it waits on a folio for
|
2013-07-03 22:01:51 +00:00
|
|
|
* which it is needed to do the write (loop masks off
|
|
|
|
* __GFP_IO|__GFP_FS for this reason); but more thought
|
|
|
|
* would probably show more reasons.
|
|
|
|
*
|
2022-05-13 03:23:02 +00:00
|
|
|
* 3) Legacy memcg encounters a folio that already has the
|
|
|
|
* reclaim flag set. memcg does not have any dirty folio
|
2013-07-03 22:01:51 +00:00
|
|
|
* throttling so we could easily OOM just because too many
|
2022-05-13 03:23:02 +00:00
|
|
|
* folios are in writeback and there is nothing else to
|
2013-07-03 22:01:51 +00:00
|
|
|
* reclaim. Wait for the writeback to complete.
|
mm: vmscan: move dirty pages out of the way until they're flushed
We noticed a performance regression when moving hadoop workloads from
3.10 kernels to 4.0 and 4.6. This is accompanied by increased pageout
activity initiated by kswapd as well as frequent bursts of allocation
stalls and direct reclaim scans. Even lowering the dirty ratios to the
equivalent of less than 1% of memory would not eliminate the issue,
suggesting that dirty pages concentrate where the scanner is looking.
This can be traced back to recent efforts of thrash avoidance. Where
3.10 would not detect refaulting pages and continuously supply clean
cache to the inactive list, a thrashing workload on 4.0+ will detect and
activate refaulting pages right away, distilling used-once pages on the
inactive list much more effectively. This is by design, and it makes
sense for clean cache. But for the most part our workload's cache
faults are refaults and its use-once cache is from streaming writes. We
end up with most of the inactive list dirty, and we don't go after the
active cache as long as we have use-once pages around.
But waiting for writes to avoid reclaiming clean cache that *might*
refault is a bad trade-off. Even if the refaults happen, reads are
faster than writes. Before getting bogged down on writeback, reclaim
should first look at *all* cache in the system, even active cache.
To accomplish this, activate pages that are dirty or under writeback
when they reach the end of the inactive LRU. The pages are marked for
immediate reclaim, meaning they'll get moved back to the inactive LRU
tail as soon as they're written back and become reclaimable. But in the
meantime, by reducing the inactive list to only immediately reclaimable
pages, we allow the scanner to deactivate and refill the inactive list
with clean cache from the active list tail to guarantee forward
progress.
[hannes@cmpxchg.org: update comment]
Link: http://lkml.kernel.org/r/20170202191957.22872-8-hannes@cmpxchg.org
Link: http://lkml.kernel.org/r/20170123181641.23938-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-24 22:56:23 +00:00
|
|
|
*
|
2022-05-13 03:23:02 +00:00
|
|
|
* In cases 1) and 2) we activate the folios to get them out of
|
|
|
|
* the way while we continue scanning for clean folios on the
|
mm: vmscan: move dirty pages out of the way until they're flushed
We noticed a performance regression when moving hadoop workloads from
3.10 kernels to 4.0 and 4.6. This is accompanied by increased pageout
activity initiated by kswapd as well as frequent bursts of allocation
stalls and direct reclaim scans. Even lowering the dirty ratios to the
equivalent of less than 1% of memory would not eliminate the issue,
suggesting that dirty pages concentrate where the scanner is looking.
This can be traced back to recent efforts of thrash avoidance. Where
3.10 would not detect refaulting pages and continuously supply clean
cache to the inactive list, a thrashing workload on 4.0+ will detect and
activate refaulting pages right away, distilling used-once pages on the
inactive list much more effectively. This is by design, and it makes
sense for clean cache. But for the most part our workload's cache
faults are refaults and its use-once cache is from streaming writes. We
end up with most of the inactive list dirty, and we don't go after the
active cache as long as we have use-once pages around.
But waiting for writes to avoid reclaiming clean cache that *might*
refault is a bad trade-off. Even if the refaults happen, reads are
faster than writes. Before getting bogged down on writeback, reclaim
should first look at *all* cache in the system, even active cache.
To accomplish this, activate pages that are dirty or under writeback
when they reach the end of the inactive LRU. The pages are marked for
immediate reclaim, meaning they'll get moved back to the inactive LRU
tail as soon as they're written back and become reclaimable. But in the
meantime, by reducing the inactive list to only immediately reclaimable
pages, we allow the scanner to deactivate and refill the inactive list
with clean cache from the active list tail to guarantee forward
progress.
[hannes@cmpxchg.org: update comment]
Link: http://lkml.kernel.org/r/20170202191957.22872-8-hannes@cmpxchg.org
Link: http://lkml.kernel.org/r/20170123181641.23938-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-24 22:56:23 +00:00
|
|
|
* inactive list and refilling from the active list. The
|
|
|
|
* observation here is that waiting for disk writes is more
|
|
|
|
* expensive than potentially causing reloads down the line.
|
|
|
|
* Since they're marked for immediate reclaim, they won't put
|
|
|
|
* memory pressure on the cache working set any longer than it
|
|
|
|
* takes to write them to disk.
|
2013-07-03 22:01:51 +00:00
|
|
|
*/
|
2022-05-13 03:23:02 +00:00
|
|
|
if (folio_test_writeback(folio)) {
|
2013-07-03 22:01:51 +00:00
|
|
|
/* Case 1 above */
|
|
|
|
if (current_is_kswapd() &&
|
2022-05-13 03:23:02 +00:00
|
|
|
folio_test_reclaim(folio) &&
|
2016-07-28 22:45:31 +00:00
|
|
|
test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
|
2022-01-17 19:46:55 +00:00
|
|
|
stat->nr_immediate += nr_pages;
|
mm: vmscan: move dirty pages out of the way until they're flushed
We noticed a performance regression when moving hadoop workloads from
3.10 kernels to 4.0 and 4.6. This is accompanied by increased pageout
activity initiated by kswapd as well as frequent bursts of allocation
stalls and direct reclaim scans. Even lowering the dirty ratios to the
equivalent of less than 1% of memory would not eliminate the issue,
suggesting that dirty pages concentrate where the scanner is looking.
This can be traced back to recent efforts of thrash avoidance. Where
3.10 would not detect refaulting pages and continuously supply clean
cache to the inactive list, a thrashing workload on 4.0+ will detect and
activate refaulting pages right away, distilling used-once pages on the
inactive list much more effectively. This is by design, and it makes
sense for clean cache. But for the most part our workload's cache
faults are refaults and its use-once cache is from streaming writes. We
end up with most of the inactive list dirty, and we don't go after the
active cache as long as we have use-once pages around.
But waiting for writes to avoid reclaiming clean cache that *might*
refault is a bad trade-off. Even if the refaults happen, reads are
faster than writes. Before getting bogged down on writeback, reclaim
should first look at *all* cache in the system, even active cache.
To accomplish this, activate pages that are dirty or under writeback
when they reach the end of the inactive LRU. The pages are marked for
immediate reclaim, meaning they'll get moved back to the inactive LRU
tail as soon as they're written back and become reclaimable. But in the
meantime, by reducing the inactive list to only immediately reclaimable
pages, we allow the scanner to deactivate and refill the inactive list
with clean cache from the active list tail to guarantee forward
progress.
[hannes@cmpxchg.org: update comment]
Link: http://lkml.kernel.org/r/20170202191957.22872-8-hannes@cmpxchg.org
Link: http://lkml.kernel.org/r/20170123181641.23938-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-24 22:56:23 +00:00
|
|
|
goto activate_locked;
|
2013-07-03 22:01:51 +00:00
|
|
|
|
|
|
|
/* Case 2 above */
|
2019-12-01 01:55:40 +00:00
|
|
|
} else if (writeback_throttling_sane(sc) ||
|
2022-05-13 03:23:02 +00:00
|
|
|
!folio_test_reclaim(folio) ||
|
2022-05-13 03:23:03 +00:00
|
|
|
!may_enter_fs(folio, sc->gfp_mask)) {
|
memcg: further prevent OOM with too many dirty pages
The may_enter_fs test turns out to be too restrictive: though I saw no
problem with it when testing on 3.5-rc6, it very soon OOMed when I tested
on 3.5-rc6-mm1. I don't know what the difference there is, perhaps I just
slightly changed the way I started off the testing: dd if=/dev/zero
of=/mnt/temp bs=1M count=1024; rm -f /mnt/temp; sync repeatedly, in 20M
memory.limit_in_bytes cgroup to ext4 on USB stick.
ext4 (and gfs2 and xfs) turn out to allocate new pages for writing with
AOP_FLAG_NOFS: that seems a little worrying, and it's unclear to me why
the transaction needs to be started even before allocating pagecache
memory. But it may not be worth worrying about these days: if direct
reclaim avoids FS writeback, does __GFP_FS now mean anything?
Anyway, we insisted on the may_enter_fs test to avoid hangs with the loop
device; but since that also masks off __GFP_IO, we can test for __GFP_IO
directly, ignoring may_enter_fs and __GFP_FS.
But even so, the test still OOMs sometimes: when originally testing on
3.5-rc6, it OOMed about one time in five or ten; when testing just now on
3.5-rc6-mm1, it OOMed on the first iteration.
This residual problem comes from an accumulation of pages under ordinary
writeback, not marked PageReclaim, so rightly not causing the memcg check
to wait on their writeback: these too can prevent shrink_page_list() from
freeing any pages, so many times that memcg reclaim fails and OOMs.
Deal with these in the same way as direct reclaim now deals with dirty FS
pages: mark them PageReclaim. It is appropriate to rotate these to tail
of list when writepage completes, but more importantly, the PageReclaim
flag makes memcg reclaim wait on them if encountered again. Increment
NR_VMSCAN_IMMEDIATE? That's arguable: I chose not.
Setting PageReclaim here may occasionally race with end_page_writeback()
clearing it: lru_deactivate_fn() already faced the same race, and
correctly concluded that the window is small and the issue non-critical.
With these changes, the test runs indefinitely without OOMing on ext4,
ext3 and ext2: I'll move on to test with other filesystems later.
Trivia: invert conditions for a clearer block without an else, and goto
keep_locked to do the unlock_page.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujtisu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Ying Han <yinghan@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Fengguang Wu <fengguang.wu@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:45:59 +00:00
|
|
|
/*
|
2022-05-13 03:23:02 +00:00
|
|
|
* This is slightly racy -
|
2022-05-13 03:23:03 +00:00
|
|
|
* folio_end_writeback() might have
|
|
|
|
* just cleared the reclaim flag, then
|
|
|
|
* setting the reclaim flag here ends up
|
|
|
|
* interpreted as the readahead flag - but
|
|
|
|
* that does not matter enough to care.
|
|
|
|
* What we do want is for this folio to
|
|
|
|
* have the reclaim flag set next time
|
|
|
|
* memcg reclaim reaches the tests above,
|
|
|
|
* so it will then wait for writeback to
|
|
|
|
* avoid OOM; and it's also appropriate
|
2022-05-13 03:23:02 +00:00
|
|
|
* in global reclaim.
|
memcg: further prevent OOM with too many dirty pages
The may_enter_fs test turns out to be too restrictive: though I saw no
problem with it when testing on 3.5-rc6, it very soon OOMed when I tested
on 3.5-rc6-mm1. I don't know what the difference there is, perhaps I just
slightly changed the way I started off the testing: dd if=/dev/zero
of=/mnt/temp bs=1M count=1024; rm -f /mnt/temp; sync repeatedly, in 20M
memory.limit_in_bytes cgroup to ext4 on USB stick.
ext4 (and gfs2 and xfs) turn out to allocate new pages for writing with
AOP_FLAG_NOFS: that seems a little worrying, and it's unclear to me why
the transaction needs to be started even before allocating pagecache
memory. But it may not be worth worrying about these days: if direct
reclaim avoids FS writeback, does __GFP_FS now mean anything?
Anyway, we insisted on the may_enter_fs test to avoid hangs with the loop
device; but since that also masks off __GFP_IO, we can test for __GFP_IO
directly, ignoring may_enter_fs and __GFP_FS.
But even so, the test still OOMs sometimes: when originally testing on
3.5-rc6, it OOMed about one time in five or ten; when testing just now on
3.5-rc6-mm1, it OOMed on the first iteration.
This residual problem comes from an accumulation of pages under ordinary
writeback, not marked PageReclaim, so rightly not causing the memcg check
to wait on their writeback: these too can prevent shrink_page_list() from
freeing any pages, so many times that memcg reclaim fails and OOMs.
Deal with these in the same way as direct reclaim now deals with dirty FS
pages: mark them PageReclaim. It is appropriate to rotate these to tail
of list when writepage completes, but more importantly, the PageReclaim
flag makes memcg reclaim wait on them if encountered again. Increment
NR_VMSCAN_IMMEDIATE? That's arguable: I chose not.
Setting PageReclaim here may occasionally race with end_page_writeback()
clearing it: lru_deactivate_fn() already faced the same race, and
correctly concluded that the window is small and the issue non-critical.
With these changes, the test runs indefinitely without OOMing on ext4,
ext3 and ext2: I'll move on to test with other filesystems later.
Trivia: invert conditions for a clearer block without an else, and goto
keep_locked to do the unlock_page.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujtisu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Ying Han <yinghan@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Fengguang Wu <fengguang.wu@intel.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:45:59 +00:00
|
|
|
*/
|
2022-05-13 03:23:02 +00:00
|
|
|
folio_set_reclaim(folio);
|
2022-01-17 19:46:55 +00:00
|
|
|
stat->nr_writeback += nr_pages;
|
mm: vmscan: move dirty pages out of the way until they're flushed
We noticed a performance regression when moving hadoop workloads from
3.10 kernels to 4.0 and 4.6. This is accompanied by increased pageout
activity initiated by kswapd as well as frequent bursts of allocation
stalls and direct reclaim scans. Even lowering the dirty ratios to the
equivalent of less than 1% of memory would not eliminate the issue,
suggesting that dirty pages concentrate where the scanner is looking.
This can be traced back to recent efforts of thrash avoidance. Where
3.10 would not detect refaulting pages and continuously supply clean
cache to the inactive list, a thrashing workload on 4.0+ will detect and
activate refaulting pages right away, distilling used-once pages on the
inactive list much more effectively. This is by design, and it makes
sense for clean cache. But for the most part our workload's cache
faults are refaults and its use-once cache is from streaming writes. We
end up with most of the inactive list dirty, and we don't go after the
active cache as long as we have use-once pages around.
But waiting for writes to avoid reclaiming clean cache that *might*
refault is a bad trade-off. Even if the refaults happen, reads are
faster than writes. Before getting bogged down on writeback, reclaim
should first look at *all* cache in the system, even active cache.
To accomplish this, activate pages that are dirty or under writeback
when they reach the end of the inactive LRU. The pages are marked for
immediate reclaim, meaning they'll get moved back to the inactive LRU
tail as soon as they're written back and become reclaimable. But in the
meantime, by reducing the inactive list to only immediately reclaimable
pages, we allow the scanner to deactivate and refill the inactive list
with clean cache from the active list tail to guarantee forward
progress.
[hannes@cmpxchg.org: update comment]
Link: http://lkml.kernel.org/r/20170202191957.22872-8-hannes@cmpxchg.org
Link: http://lkml.kernel.org/r/20170123181641.23938-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-24 22:56:23 +00:00
|
|
|
goto activate_locked;
|
2013-07-03 22:01:51 +00:00
|
|
|
|
|
|
|
/* Case 3 above */
|
|
|
|
} else {
|
2022-05-13 03:23:02 +00:00
|
|
|
folio_unlock(folio);
|
|
|
|
folio_wait_writeback(folio);
|
|
|
|
/* then go back and try same folio again */
|
2022-09-02 19:45:57 +00:00
|
|
|
list_add_tail(&folio->lru, folio_list);
|
2015-09-08 22:03:46 +00:00
|
|
|
continue;
|
memcg: prevent OOM with too many dirty pages
The current implementation of dirty pages throttling is not memcg aware
which makes it easy to have memcg LRUs full of dirty pages. Without
throttling, these LRUs can be scanned faster than the rate of writeback,
leading to memcg OOM conditions when the hard limit is small.
This patch fixes the problem by throttling the allocating process
(possibly a writer) during the hard limit reclaim by waiting on
PageReclaim pages. We are waiting only for PageReclaim pages because
those are the pages that made one full round over LRU and that means that
the writeback is much slower than scanning.
The solution is far from being ideal - long term solution is memcg aware
dirty throttling - but it is meant to be a band aid until we have a real
fix. We are seeing this happening during nightly backups which are placed
into containers to prevent from eviction of the real working set.
The change affects only memcg reclaim and only when we encounter
PageReclaim pages which is a signal that the reclaim doesn't catch up on
with the writers so somebody should be throttled. This could be
potentially unfair because it could be somebody else from the group who
gets throttled on behalf of the writer but as writers need to allocate as
well and they allocate in higher rate the probability that only innocent
processes would be penalized is not that high.
I have tested this change by a simple dd copying /dev/zero to tmpfs or
ext3 running under small memcg (1G copy under 5M, 60M, 300M and 2G
containers) and dd got killed by OOM killer every time. With the patch I
could run the dd with the same size under 5M controller without any OOM.
The issue is more visible with slower devices for output.
* With the patch
================
* tmpfs size=2G
---------------
$ vim cgroup_cache_oom_test.sh
$ ./cgroup_cache_oom_test.sh 5M
using Limit 5M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 30.4049 s, 34.5 MB/s
$ ./cgroup_cache_oom_test.sh 60M
using Limit 60M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 31.4561 s, 33.3 MB/s
$ ./cgroup_cache_oom_test.sh 300M
using Limit 300M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 20.4618 s, 51.2 MB/s
$ ./cgroup_cache_oom_test.sh 2G
using Limit 2G for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 1.42172 s, 738 MB/s
* ext3
------
$ ./cgroup_cache_oom_test.sh 5M
using Limit 5M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 27.9547 s, 37.5 MB/s
$ ./cgroup_cache_oom_test.sh 60M
using Limit 60M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 30.3221 s, 34.6 MB/s
$ ./cgroup_cache_oom_test.sh 300M
using Limit 300M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 24.5764 s, 42.7 MB/s
$ ./cgroup_cache_oom_test.sh 2G
using Limit 2G for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 3.35828 s, 312 MB/s
* Without the patch
===================
* tmpfs size=2G
---------------
$ ./cgroup_cache_oom_test.sh 5M
using Limit 5M for group
./cgroup_cache_oom_test.sh: line 46: 4668 Killed dd if=/dev/zero of=$OUT/zero bs=1M count=$count
$ ./cgroup_cache_oom_test.sh 60M
using Limit 60M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 25.4989 s, 41.1 MB/s
$ ./cgroup_cache_oom_test.sh 300M
using Limit 300M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 24.3928 s, 43.0 MB/s
$ ./cgroup_cache_oom_test.sh 2G
using Limit 2G for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 1.49797 s, 700 MB/s
* ext3
------
$ ./cgroup_cache_oom_test.sh 5M
using Limit 5M for group
./cgroup_cache_oom_test.sh: line 46: 4689 Killed dd if=/dev/zero of=$OUT/zero bs=1M count=$count
$ ./cgroup_cache_oom_test.sh 60M
using Limit 60M for group
./cgroup_cache_oom_test.sh: line 46: 4692 Killed dd if=/dev/zero of=$OUT/zero bs=1M count=$count
$ ./cgroup_cache_oom_test.sh 300M
using Limit 300M for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 20.248 s, 51.8 MB/s
$ ./cgroup_cache_oom_test.sh 2G
using Limit 2G for group
1000+0 records in
1000+0 records out
1048576000 bytes (1.0 GB) copied, 2.85201 s, 368 MB/s
[akpm@linux-foundation.org: tweak changelog, reordered the test to optimize for CONFIG_CGROUP_MEM_RES_CTLR=n]
[hughd@google.com: fix deadlock with loop driver]
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujtisu.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Ying Han <yinghan@google.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Reviewed-by: Mel Gorman <mgorman@suse.de>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:45:55 +00:00
|
|
|
}
|
2007-08-22 21:01:26 +00:00
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2019-09-25 23:49:11 +00:00
|
|
|
if (!ignore_references)
|
2022-02-15 18:44:40 +00:00
|
|
|
references = folio_check_references(folio, sc);
|
2012-10-08 23:31:55 +00:00
|
|
|
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
switch (references) {
|
2022-09-02 19:45:57 +00:00
|
|
|
case FOLIOREF_ACTIVATE:
|
2005-04-16 22:20:36 +00:00
|
|
|
goto activate_locked;
|
2022-09-02 19:45:57 +00:00
|
|
|
case FOLIOREF_KEEP:
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
stat->nr_ref_keep += nr_pages;
|
2010-03-05 21:42:22 +00:00
|
|
|
goto keep_locked;
|
2022-09-02 19:45:57 +00:00
|
|
|
case FOLIOREF_RECLAIM:
|
|
|
|
case FOLIOREF_RECLAIM_CLEAN:
|
2022-05-13 03:23:03 +00:00
|
|
|
; /* try to reclaim the folio below */
|
vmscan: factor out page reference checks
The used-once mapped file page detection patchset.
It is meant to help workloads with large amounts of shortly used file
mappings, like rtorrent hashing a file or git when dealing with loose
objects (git gc on a bigger site?).
Right now, the VM activates referenced mapped file pages on first
encounter on the inactive list and it takes a full memory cycle to
reclaim them again. When those pages dominate memory, the system
no longer has a meaningful notion of 'working set' and is required
to give up the active list to make reclaim progress. Obviously,
this results in rather bad scanning latencies and the wrong pages
being reclaimed.
This patch makes the VM be more careful about activating mapped file
pages in the first place. The minimum granted lifetime without
another memory access becomes an inactive list cycle instead of the
full memory cycle, which is more natural given the mentioned loads.
This test resembles a hashing rtorrent process. Sequentially, 32MB
chunks of a file are mapped into memory, hashed (sha1) and unmapped
again. While this happens, every 5 seconds a process is launched and
its execution time taken:
python2.4 -c 'import pydoc'
old: max=2.31s mean=1.26s (0.34)
new: max=1.25s mean=0.32s (0.32)
find /etc -type f
old: max=2.52s mean=1.44s (0.43)
new: max=1.92s mean=0.12s (0.17)
vim -c ':quit'
old: max=6.14s mean=4.03s (0.49)
new: max=3.48s mean=2.41s (0.25)
mplayer --help
old: max=8.08s mean=5.74s (1.02)
new: max=3.79s mean=1.32s (0.81)
overall hash time (stdev):
old: time=1192.30 (12.85) thruput=25.78mb/s (0.27)
new: time=1060.27 (32.58) thruput=29.02mb/s (0.88) (-11%)
I also tested kernbench with regular IO streaming in the background to
see whether the delayed activation of frequently used mapped file
pages had a negative impact on performance in the presence of pressure
on the inactive list. The patch made no significant difference in
timing, neither for kernbench nor for the streaming IO throughput.
The first patch submission raised concerns about the cost of the extra
faults for actually activated pages on machines that have no hardware
support for young page table entries.
I created an artificial worst case scenario on an ARM machine with
around 300MHz and 64MB of memory to figure out the dimensions
involved. The test would mmap a file of 20MB, then
1. touch all its pages to fault them in
2. force one full scan cycle on the inactive file LRU
-- old: mapping pages activated
-- new: mapping pages inactive
3. touch the mapping pages again
-- old and new: fault exceptions to set the young bits
4. force another full scan cycle on the inactive file LRU
5. touch the mapping pages one last time
-- new: fault exceptions to set the young bits
The test showed an overall increase of 6% in time over 100 iterations
of the above (old: ~212sec, new: ~225sec). 13 secs total overhead /
(100 * 5k pages), ignoring the execution time of the test itself,
makes for about 25us overhead for every page that gets actually
activated. Note:
1. File mapping the size of one third of main memory, _completely_
in active use across memory pressure - i.e., most pages referenced
within one LRU cycle. This should be rare to non-existant,
especially on such embedded setups.
2. Many huge activation batches. Those batches only occur when the
working set fluctuates. If it changes completely between every full
LRU cycle, you have problematic reclaim overhead anyway.
3. Access of activated pages at maximum speed: sequential loads from
every single page without doing anything in between. In reality,
the extra faults will get distributed between actual operations on
the data.
So even if a workload manages to get the VM into the situation of
activating a third of memory in one go on such a setup, it will take
2.2 seconds instead 2.1 without the patch.
Comparing the numbers (and my user-experience over several months),
I think this change is an overall improvement to the VM.
Patch 1 is only refactoring to break up that ugly compound conditional
in shrink_page_list() and make it easy to document and add new checks
in a readable fashion.
Patch 2 gets rid of the obsolete page_mapping_inuse(). It's not
strictly related to #3, but it was in the original submission and is a
net simplification, so I kept it.
Patch 3 implements used-once detection of mapped file pages.
This patch:
Moving the big conditional into its own predicate function makes the code
a bit easier to read and allows for better commenting on the checks
one-by-one.
This is just cleaning up, no semantics should have been changed.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: OSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-03-05 21:42:19 +00:00
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2021-09-02 21:59:16 +00:00
|
|
|
/*
|
2022-05-13 03:23:03 +00:00
|
|
|
* Before reclaiming the folio, try to relocate
|
2021-09-02 21:59:16 +00:00
|
|
|
* its contents to another node.
|
|
|
|
*/
|
|
|
|
if (do_demote_pass &&
|
2022-05-13 03:23:03 +00:00
|
|
|
(thp_migration_supported() || !folio_test_large(folio))) {
|
2022-09-02 19:45:57 +00:00
|
|
|
list_add(&folio->lru, &demote_folios);
|
2022-05-13 03:23:03 +00:00
|
|
|
folio_unlock(folio);
|
2021-09-02 21:59:16 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* Anonymous process memory has backing store?
|
|
|
|
* Try to allocate it some swap space here.
|
2022-05-13 03:23:03 +00:00
|
|
|
* Lazyfree folio could be freed directly
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2022-05-13 03:23:03 +00:00
|
|
|
if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
|
|
|
|
if (!folio_test_swapcache(folio)) {
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
if (!(sc->gfp_mask & __GFP_IO))
|
|
|
|
goto keep_locked;
|
2022-02-04 19:13:31 +00:00
|
|
|
if (folio_maybe_dma_pinned(folio))
|
mm: don't put pinned pages into the swap cache
So technically there is nothing wrong with adding a pinned page to the
swap cache, but the pinning obviously means that the page can't actually
be free'd right now anyway, so it's a bit pointless.
However, the real problem is not with it being a bit pointless: the real
issue is that after we've added it to the swap cache, we'll try to unmap
the page. That will succeed, because the code in mm/rmap.c doesn't know
or care about pinned pages.
Even the unmapping isn't fatal per se, since the page will stay around
in memory due to the pinning, and we do hold the connection to it using
the swap cache. But when we then touch it next and take a page fault,
the logic in do_swap_page() will map it back into the process as a
possibly read-only page, and we'll then break the page association on
the next COW fault.
Honestly, this issue could have been fixed in any of those other places:
(a) we could refuse to unmap a pinned page (which makes conceptual
sense), or (b) we could make sure to re-map a pinned page writably in
do_swap_page(), or (c) we could just make do_wp_page() not COW the
pinned page (which was what we historically did before that "mm:
do_wp_page() simplification" commit).
But while all of them are equally valid models for breaking this chain,
not putting pinned pages into the swap cache in the first place is the
simplest one by far.
It's also the safest one: the reason why do_wp_page() was changed in the
first place was that getting the "can I re-use this page" wrong is so
fraught with errors. If you do it wrong, you end up with an incorrectly
shared page.
As a result, using "page_maybe_dma_pinned()" in either do_wp_page() or
do_swap_page() would be a serious bug since it is only a (very good)
heuristic. Re-using the page requires a hard black-and-white rule with
no room for ambiguity.
In contrast, saying "this page is very likely dma pinned, so let's not
add it to the swap cache and try to unmap it" is an obviously safe thing
to do, and if the heuristic might very rarely be a false positive, no
harm is done.
Fixes: 09854ba94c6a ("mm: do_wp_page() simplification")
Reported-and-tested-by: Martin Raiber <martin@urbackup.org>
Cc: Pavel Begunkov <asml.silence@gmail.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-01-16 23:34:57 +00:00
|
|
|
goto keep_locked;
|
2022-05-13 03:23:03 +00:00
|
|
|
if (folio_test_large(folio)) {
|
|
|
|
/* cannot split folio, skip it */
|
2022-02-04 19:13:31 +00:00
|
|
|
if (!can_split_folio(folio, NULL))
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
goto activate_locked;
|
|
|
|
/*
|
2022-05-13 03:23:03 +00:00
|
|
|
* Split folios without a PMD map right
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
* away. Chances are some or all of the
|
|
|
|
* tail pages can be freed without IO.
|
|
|
|
*/
|
2022-02-04 19:13:31 +00:00
|
|
|
if (!folio_entire_mapcount(folio) &&
|
2022-01-18 13:56:47 +00:00
|
|
|
split_folio_to_list(folio,
|
2022-09-02 19:45:57 +00:00
|
|
|
folio_list))
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
goto activate_locked;
|
|
|
|
}
|
2022-05-13 03:23:02 +00:00
|
|
|
if (!add_to_swap(folio)) {
|
|
|
|
if (!folio_test_large(folio))
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
goto activate_locked_split;
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
/* Fallback to swap normal pages */
|
2022-01-18 13:56:47 +00:00
|
|
|
if (split_folio_to_list(folio,
|
2022-09-02 19:45:57 +00:00
|
|
|
folio_list))
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
goto activate_locked;
|
mm, THP, swap: add THP swapping out fallback counting
When swapping out THP (Transparent Huge Page), instead of swapping out
the THP as a whole, sometimes we have to fallback to split the THP into
normal pages before swapping, because no free swap clusters are
available, or cgroup limit is exceeded, etc. To count the number of the
fallback, a new VM event THP_SWPOUT_FALLBACK is added, and counted when
we fallback to split the THP.
Link: http://lkml.kernel.org/r/20170724051840.2309-13-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:52 +00:00
|
|
|
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
|
|
|
|
count_vm_event(THP_SWPOUT_FALLBACK);
|
|
|
|
#endif
|
2022-05-13 03:23:02 +00:00
|
|
|
if (!add_to_swap(folio))
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
goto activate_locked_split;
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
}
|
|
|
|
}
|
2022-05-13 03:23:03 +00:00
|
|
|
} else if (folio_test_swapbacked(folio) &&
|
|
|
|
folio_test_large(folio)) {
|
|
|
|
/* Split shmem folio */
|
2022-09-02 19:45:57 +00:00
|
|
|
if (split_folio_to_list(folio, folio_list))
|
2016-07-26 22:25:56 +00:00
|
|
|
goto keep_locked;
|
mm: vmscan: stall page reclaim and writeback pages based on dirty/writepage pages encountered
Further testing of the "Reduce system disruption due to kswapd"
discovered a few problems. First and foremost, it's possible for pages
under writeback to be freed which will lead to badness. Second, as
pages were not being swapped the file LRU was being scanned faster and
clean file pages were being reclaimed. In some cases this results in
increased read IO to re-read data from disk. Third, more pages were
being written from kswapd context which can adversly affect IO
performance. Lastly, it was observed that PageDirty pages are not
necessarily dirty on all filesystems (buffers can be clean while
PageDirty is set and ->writepage generates no IO) and not all
filesystems set PageWriteback when the page is being written (e.g.
ext3). This disconnect confuses the reclaim stalling logic. This
follow-up series is aimed at these problems.
The tests were based on three kernels
vanilla: kernel 3.9 as that is what the current mmotm uses as a baseline
mmotm-20130522 is mmotm as of 22nd May with "Reduce system disruption due to
kswapd" applied on top as per what should be in Andrew's tree
right now
lessdisrupt-v7r10 is this follow-up series on top of the mmotm kernel
The first test used memcached+memcachetest while some background IO was
in progress as implemented by the parallel IO tests implement in MM
Tests. memcachetest benchmarks how many operations/second memcached can
service. It starts with no background IO on a freshly created ext4
filesystem and then re-runs the test with larger amounts of IO in the
background to roughly simulate a large copy in progress. The
expectation is that the IO should have little or no impact on
memcachetest which is running entirely in memory.
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23117.00 ( 0.00%) 22780.00 ( -1.46%) 22763.00 ( -1.53%)
Ops memcachetest-715M 23774.00 ( 0.00%) 23299.00 ( -2.00%) 22934.00 ( -3.53%)
Ops memcachetest-2385M 4208.00 ( 0.00%) 24154.00 (474.00%) 23765.00 (464.76%)
Ops memcachetest-4055M 4104.00 ( 0.00%) 25130.00 (512.33%) 24614.00 (499.76%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%) 6.00 ( 50.00%)
Ops io-duration-2385M 116.00 ( 0.00%) 21.00 ( 81.90%) 21.00 ( 81.90%)
Ops io-duration-4055M 160.00 ( 0.00%) 36.00 ( 77.50%) 35.00 ( 78.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140138.00 ( 0.00%) 18.00 ( 99.99%) 18.00 ( 99.99%)
Ops swaptotal-2385M 385682.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 418029.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 144.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 134227.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 125618.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1536429.00 ( 0.00%) 1531632.00 ( 0.31%) 1533541.00 ( 0.19%)
Ops minorfaults-715M 1786996.00 ( 0.00%) 1612148.00 ( 9.78%) 1608832.00 ( 9.97%)
Ops minorfaults-2385M 1757952.00 ( 0.00%) 1614874.00 ( 8.14%) 1613541.00 ( 8.21%)
Ops minorfaults-4055M 1774460.00 ( 0.00%) 1633400.00 ( 7.95%) 1630881.00 ( 8.09%)
Ops majorfaults-0M 1.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 184.00 ( 0.00%) 167.00 ( 9.24%) 166.00 ( 9.78%)
Ops majorfaults-2385M 24444.00 ( 0.00%) 155.00 ( 99.37%) 93.00 ( 99.62%)
Ops majorfaults-4055M 21357.00 ( 0.00%) 147.00 ( 99.31%) 134.00 ( 99.37%)
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
23K/sec to just over 4K/second when there is 2385M of IO going
on in the background. With current mmotm, there is no collapse
in performance and with this follow-up series there is little
change.
swaptotal is the total amount of swap traffic. With mmotm and the follow-up
series, the total amount of swapping is much reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 11160152 10706748 10622316
Major Faults 46305 755 678
Swap Ins 260249 0 0
Swap Outs 683860 18 18
Direct pages scanned 0 678 2520
Kswapd pages scanned 6046108 8814900 1639279
Kswapd pages reclaimed 1081954 1172267 1094635
Direct pages reclaimed 0 566 2304
Kswapd efficiency 17% 13% 66%
Kswapd velocity 5217.560 7618.953 1414.879
Direct efficiency 100% 83% 91%
Direct velocity 0.000 0.586 2.175
Percentage direct scans 0% 0% 0%
Zone normal velocity 5105.086 6824.681 671.158
Zone dma32 velocity 112.473 794.858 745.896
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 1929612.000 6861768.000 32821.000
Page writes file 1245752 6861750 32803
Page writes anon 683860 18 18
Page reclaim immediate 7484 40 239
Sector Reads 1130320 93996 86900
Sector Writes 13508052 10823500 11804436
Page rescued immediate 0 0 0
Slabs scanned 33536 27136 18560
Direct inode steals 0 0 0
Kswapd inode steals 8641 1035 0
Kswapd skipped wait 0 0 0
THP fault alloc 8 37 33
THP collapse alloc 508 552 515
THP splits 24 1 1
THP fault fallback 0 0 0
THP collapse fail 0 0 0
There are a number of observations to make here
1. Swap outs are almost eliminated. Swap ins are 0 indicating that the
pages swapped were really unused anonymous pages. Related to that,
major faults are much reduced.
2. kswapd efficiency was impacted by the initial series but with these
follow-up patches, the efficiency is now at 66% indicating that far
fewer pages were skipped during scanning due to dirty or writeback
pages.
3. kswapd velocity is reduced indicating that fewer pages are being scanned
with the follow-up series as kswapd now stalls when the tail of the
LRU queue is full of unqueued dirty pages. The stall gives flushers a
chance to catch-up so kswapd can reclaim clean pages when it wakes
4. In light of Zlatko's recent reports about zone scanning imbalances,
mmtests now reports scanning velocity on a per-zone basis. With mainline,
you can see that the scanning activity is dominated by the Normal
zone with over 45 times more scanning in Normal than the DMA32 zone.
With the series currently in mmotm, the ratio is slightly better but it
is still the case that the bulk of scanning is in the highest zone. With
this follow-up series, the ratio of scanning between the Normal and
DMA32 zone is roughly equal.
5. As Dave Chinner observed, the current patches in mmotm increased the
number of pages written from kswapd context which is expected to adversly
impact IO performance. With the follow-up patches, far fewer pages are
written from kswapd context than the mainline kernel
6. With the series in mmotm, fewer inodes were reclaimed by kswapd. With
the follow-up series, there is less slab shrinking activity and no inodes
were reclaimed.
7. Note that "Sectors Read" is drastically reduced implying that the source
data being used for the IO is not being aggressively discarded due to
page reclaim skipping over dirty pages and reclaiming clean pages. Note
that the reducion in reads could also be due to inode data not being
re-read from disk after a slab shrink.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 166.99 32.09 33.44
Mean sda-await 853.64 192.76 185.43
Mean sda-r_await 6.31 9.24 5.97
Mean sda-w_await 2992.81 202.65 192.43
Max sda-avgqz 1409.91 718.75 698.98
Max sda-await 6665.74 3538.00 3124.23
Max sda-r_await 58.96 111.95 58.00
Max sda-w_await 28458.94 3977.29 3148.61
In light of the changes in writes from reclaim context, the number of
reads and Dave Chinner's concerns about IO performance I took a closer
look at the IO stats for the test disk. Few observations
1. The average queue size is reduced by the initial series and roughly
the same with this follow up.
2. Average wait times for writes are reduced and as the IO
is completing faster it at least implies that the gain is because
flushers are writing the files efficiently instead of page reclaim
getting in the way.
3. The reduction in maximum write latency is staggering. 28 seconds down
to 3 seconds.
Jan Kara asked how NFS is affected by all of this. Unstable pages can
be taken into account as one of the patches in the series shows but it
is still the case that filesystems with unusual handling of dirty or
writeback could still be treated better.
Tests like postmark, fsmark and largedd showed up nothing useful. On my test
setup, pages are simply not being written back from reclaim context with or
without the patches and there are no changes in performance. My test setup
probably is just not strong enough network-wise to be really interesting.
I ran a longer-lived memcached test with IO going to NFS instead of a local disk
parallelio
3.9.0 3.9.0 3.9.0
vanilla mm1-mmotm-20130522 mm1-lessdisrupt-v7r10
Ops memcachetest-0M 23323.00 ( 0.00%) 23241.00 ( -0.35%) 23321.00 ( -0.01%)
Ops memcachetest-715M 25526.00 ( 0.00%) 24763.00 ( -2.99%) 23242.00 ( -8.95%)
Ops memcachetest-2385M 8814.00 ( 0.00%) 26924.00 (205.47%) 23521.00 (166.86%)
Ops memcachetest-4055M 5835.00 ( 0.00%) 26827.00 (359.76%) 25560.00 (338.05%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 65.00 ( 0.00%) 71.00 ( -9.23%) 11.00 ( 83.08%)
Ops io-duration-2385M 129.00 ( 0.00%) 94.00 ( 27.13%) 53.00 ( 58.91%)
Ops io-duration-4055M 301.00 ( 0.00%) 100.00 ( 66.78%) 108.00 ( 64.12%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 14394.00 ( 0.00%) 949.00 ( 93.41%) 63.00 ( 99.56%)
Ops swaptotal-2385M 401483.00 ( 0.00%) 24437.00 ( 93.91%) 30118.00 ( 92.50%)
Ops swaptotal-4055M 554123.00 ( 0.00%) 35688.00 ( 93.56%) 63082.00 ( 88.62%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 4522.00 ( 0.00%) 560.00 ( 87.62%) 63.00 ( 98.61%)
Ops swapin-2385M 169861.00 ( 0.00%) 5026.00 ( 97.04%) 13917.00 ( 91.81%)
Ops swapin-4055M 192374.00 ( 0.00%) 10056.00 ( 94.77%) 25729.00 ( 86.63%)
Ops minorfaults-0M 1445969.00 ( 0.00%) 1520878.00 ( -5.18%) 1454024.00 ( -0.56%)
Ops minorfaults-715M 1557288.00 ( 0.00%) 1528482.00 ( 1.85%) 1535776.00 ( 1.38%)
Ops minorfaults-2385M 1692896.00 ( 0.00%) 1570523.00 ( 7.23%) 1559622.00 ( 7.87%)
Ops minorfaults-4055M 1654985.00 ( 0.00%) 1581456.00 ( 4.44%) 1596713.00 ( 3.52%)
Ops majorfaults-0M 0.00 ( 0.00%) 1.00 (-99.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 763.00 ( 0.00%) 265.00 ( 65.27%) 75.00 ( 90.17%)
Ops majorfaults-2385M 23861.00 ( 0.00%) 894.00 ( 96.25%) 2189.00 ( 90.83%)
Ops majorfaults-4055M 27210.00 ( 0.00%) 1569.00 ( 94.23%) 4088.00 ( 84.98%)
1. Performance does not collapse due to IO which is good. IO is also completing
faster. Note with mmotm, IO completes in a third of the time and faster again
with this series applied
2. Swapping is reduced, although not eliminated. The figures for the follow-up
look bad but it does vary a bit as the stalling is not perfect for nfs
or filesystems like ext3 with unusual handling of dirty and writeback
pages
3. There are swapins, particularly with larger amounts of IO indicating
that active pages are being reclaimed. However, the number of much
reduced.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Minor Faults 36339175 35025445 35219699
Major Faults 310964 27108 51887
Swap Ins 2176399 173069 333316
Swap Outs 3344050 357228 504824
Direct pages scanned 8972 77283 43242
Kswapd pages scanned 20899983 8939566 14772851
Kswapd pages reclaimed 6193156 5172605 5231026
Direct pages reclaimed 8450 73802 39514
Kswapd efficiency 29% 57% 35%
Kswapd velocity 3929.743 1847.499 3058.840
Direct efficiency 94% 95% 91%
Direct velocity 1.687 15.972 8.954
Percentage direct scans 0% 0% 0%
Zone normal velocity 3721.907 939.103 2185.142
Zone dma32 velocity 209.522 924.368 882.651
Zone dma velocity 0.000 0.000 0.000
Page writes by reclaim 4082185.000 526319.000 537114.000
Page writes file 738135 169091 32290
Page writes anon 3344050 357228 504824
Page reclaim immediate 9524 170 5595843
Sector Reads 8909900 861192 1483680
Sector Writes 13428980 1488744 2076800
Page rescued immediate 0 0 0
Slabs scanned 38016 31744 28672
Direct inode steals 0 0 0
Kswapd inode steals 424 0 0
Kswapd skipped wait 0 0 0
THP fault alloc 14 15 119
THP collapse alloc 1767 1569 1618
THP splits 30 29 25
THP fault fallback 0 0 0
THP collapse fail 8 5 0
Compaction stalls 17 41 100
Compaction success 7 31 95
Compaction failures 10 10 5
Page migrate success 7083 22157 62217
Page migrate failure 0 0 0
Compaction pages isolated 14847 48758 135830
Compaction migrate scanned 18328 48398 138929
Compaction free scanned 2000255 355827 1720269
Compaction cost 7 24 68
I guess the main takeaway again is the much reduced page writes
from reclaim context and reduced reads.
3.9.0 3.9.0 3.9.0
vanillamm1-mmotm-20130522mm1-lessdisrupt-v7r10
Mean sda-avgqz 23.58 0.35 0.44
Mean sda-await 133.47 15.72 15.46
Mean sda-r_await 4.72 4.69 3.95
Mean sda-w_await 507.69 28.40 33.68
Max sda-avgqz 680.60 12.25 23.14
Max sda-await 3958.89 221.83 286.22
Max sda-r_await 63.86 61.23 67.29
Max sda-w_await 11710.38 883.57 1767.28
And as before, write wait times are much reduced.
This patch:
The patch "mm: vmscan: Have kswapd writeback pages based on dirty pages
encountered, not priority" decides whether to writeback pages from reclaim
context based on the number of dirty pages encountered. This situation is
flagged too easily and flushers are not given the chance to catch up
resulting in more pages being written from reclaim context and potentially
impacting IO performance. The check for PageWriteback is also misplaced
as it happens within a PageDirty check which is nonsense as the dirty may
have been cleared for IO. The accounting is updated very late and pages
that are already under writeback, were reactivated, could not unmapped or
could not be released are all missed. Similarly, a page is considered
congested for reasons other than being congested and pages that cannot be
written out in the correct context are skipped. Finally, it considers
stalling and writing back filesystem pages due to encountering dirty
anonymous pages at the tail of the LRU which is dumb.
This patch causes kswapd to begin writing filesystem pages from reclaim
context only if page reclaim found that all filesystem pages at the tail
of the LRU were unqueued dirty pages. Before it starts writing filesystem
pages, it will stall to give flushers a chance to catch up. The decision
on whether wait_iff_congested is also now determined by dirty filesystem
pages only. Congested pages are based on whether the underlying BDI is
congested regardless of the context of the reclaiming process.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Cc: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:57 +00:00
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
/*
|
2022-05-13 03:23:03 +00:00
|
|
|
* If the folio was split above, the tail pages will make
|
|
|
|
* their own pass through this function and be accounted
|
|
|
|
* then.
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
*/
|
2022-05-13 03:23:03 +00:00
|
|
|
if ((nr_pages > 1) && !folio_test_large(folio)) {
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
sc->nr_scanned -= (nr_pages - 1);
|
|
|
|
nr_pages = 1;
|
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2022-05-13 03:23:01 +00:00
|
|
|
* The folio is mapped into the page tables of one or more
|
2005-04-16 22:20:36 +00:00
|
|
|
* processes. Try to unmap it here.
|
|
|
|
*/
|
2022-05-13 03:23:01 +00:00
|
|
|
if (folio_mapped(folio)) {
|
2020-12-15 03:06:39 +00:00
|
|
|
enum ttu_flags flags = TTU_BATCH_FLUSH;
|
2022-05-13 03:23:01 +00:00
|
|
|
bool was_swapbacked = folio_test_swapbacked(folio);
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
|
2022-05-13 03:23:01 +00:00
|
|
|
if (folio_test_pmd_mappable(folio))
|
mm, THP, swap: delay splitting THP after swapped out
In this patch, splitting transparent huge page (THP) during swapping out
is delayed from after adding the THP into the swap cache to after
swapping out finishes. After the patch, more operations for the
anonymous THP reclaiming, such as writing the THP to the swap device,
removing the THP from the swap cache could be batched. So that the
performance of anonymous THP swapping out could be improved.
This is the second step for the THP swap support. The plan is to delay
splitting the THP step by step and avoid splitting the THP finally.
With the patchset, the swap out throughput improves 42% (from about
5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case
with 16 processes. At the same time, the IPI (reflect TLB flushing)
reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170724051840.2309-12-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c]
Cc: Vishal L Verma <vishal.l.verma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:49 +00:00
|
|
|
flags |= TTU_SPLIT_HUGE_PMD;
|
mm/vmscan: count layzfree pages and fix nr_isolated_* mismatch
Fix an nr_isolate_* mismatch problem between cma and dirty lazyfree pages.
If try_to_unmap_one is used for reclaim and it detects a dirty lazyfree
page, then the lazyfree page is changed to a normal anon page having
SwapBacked by commit 802a3a92ad7a ("mm: reclaim MADV_FREE pages"). Even
with the change, reclaim context correctly counts isolated files because
it uses is_file_lru to distinguish file. And the change to anon is not
happened if try_to_unmap_one is used for migration. So migration context
like compaction also correctly counts isolated files even though it uses
page_is_file_lru insted of is_file_lru. Recently page_is_file_cache was
renamed to page_is_file_lru by commit 9de4f22a60f7 ("mm: code cleanup for
MADV_FREE").
But the nr_isolate_* mismatch problem happens on cma alloc. There is
reclaim_clean_pages_from_list which is being used only by cma. It was
introduced by commit 02c6de8d757c ("mm: cma: discard clean pages during
contiguous allocation instead of migration") to reclaim clean file pages
without migration. The cma alloc uses both reclaim_clean_pages_from_list
and migrate_pages, and it uses page_is_file_lru to count isolated files.
If there are dirty lazyfree pages allocated from cma memory region, the
pages are counted as isolated file at the beginging but are counted as
isolated anon after finished.
Mem-Info:
Node 0 active_anon:3045904kB inactive_anon:611448kB active_file:14892kB inactive_file:205636kB unevictable:10416kB isolated(anon):0kB isolated(file):37664kB mapped:630216kB dirty:384kB writeback:0kB shmem:42576kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Like log above, there were too much isolated files, 37664kB, which
triggers too_many_isolated in reclaim even when there is no actually
isolated file in system wide. It could be reproducible by running two
programs, writing on MADV_FREE page and doing cma alloc, respectively.
Although isolated anon is 0, I found that the internal value of isolated
anon was the negative value of isolated file.
Fix this by compensating the isolated count for both LRU lists. Count
non-discarded lazyfree pages in shrink_page_list, then compensate the
counted number in reclaim_clean_pages_from_list.
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200426011718.30246-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:01:15 +00:00
|
|
|
|
2022-02-15 14:28:49 +00:00
|
|
|
try_to_unmap(folio, flags);
|
2022-05-13 03:23:01 +00:00
|
|
|
if (folio_mapped(folio)) {
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
stat->nr_unmap_fail += nr_pages;
|
2022-05-13 03:23:01 +00:00
|
|
|
if (!was_swapbacked &&
|
|
|
|
folio_test_swapbacked(folio))
|
mm/vmscan: count layzfree pages and fix nr_isolated_* mismatch
Fix an nr_isolate_* mismatch problem between cma and dirty lazyfree pages.
If try_to_unmap_one is used for reclaim and it detects a dirty lazyfree
page, then the lazyfree page is changed to a normal anon page having
SwapBacked by commit 802a3a92ad7a ("mm: reclaim MADV_FREE pages"). Even
with the change, reclaim context correctly counts isolated files because
it uses is_file_lru to distinguish file. And the change to anon is not
happened if try_to_unmap_one is used for migration. So migration context
like compaction also correctly counts isolated files even though it uses
page_is_file_lru insted of is_file_lru. Recently page_is_file_cache was
renamed to page_is_file_lru by commit 9de4f22a60f7 ("mm: code cleanup for
MADV_FREE").
But the nr_isolate_* mismatch problem happens on cma alloc. There is
reclaim_clean_pages_from_list which is being used only by cma. It was
introduced by commit 02c6de8d757c ("mm: cma: discard clean pages during
contiguous allocation instead of migration") to reclaim clean file pages
without migration. The cma alloc uses both reclaim_clean_pages_from_list
and migrate_pages, and it uses page_is_file_lru to count isolated files.
If there are dirty lazyfree pages allocated from cma memory region, the
pages are counted as isolated file at the beginging but are counted as
isolated anon after finished.
Mem-Info:
Node 0 active_anon:3045904kB inactive_anon:611448kB active_file:14892kB inactive_file:205636kB unevictable:10416kB isolated(anon):0kB isolated(file):37664kB mapped:630216kB dirty:384kB writeback:0kB shmem:42576kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Like log above, there were too much isolated files, 37664kB, which
triggers too_many_isolated in reclaim even when there is no actually
isolated file in system wide. It could be reproducible by running two
programs, writing on MADV_FREE page and doing cma alloc, respectively.
Although isolated anon is 0, I found that the internal value of isolated
anon was the negative value of isolated file.
Fix this by compensating the isolated count for both LRU lists. Count
non-discarded lazyfree pages in shrink_page_list, then compensate the
counted number in reclaim_clean_pages_from_list.
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200426011718.30246-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:01:15 +00:00
|
|
|
stat->nr_lazyfree_fail += nr_pages;
|
2005-04-16 22:20:36 +00:00
|
|
|
goto activate_locked;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
mapping = folio_mapping(folio);
|
2022-05-13 03:23:02 +00:00
|
|
|
if (folio_test_dirty(folio)) {
|
mm: vmscan: do not writeback filesystem pages in direct reclaim
Testing from the XFS folk revealed that there is still too much I/O from
the end of the LRU in kswapd. Previously it was considered acceptable by
VM people for a small number of pages to be written back from reclaim with
testing generally showing about 0.3% of pages reclaimed were written back
(higher if memory was low). That writing back a small number of pages is
ok has been heavily disputed for quite some time and Dave Chinner
explained it well;
It doesn't have to be a very high number to be a problem. IO
is orders of magnitude slower than the CPU time it takes to
flush a page, so the cost of making a bad flush decision is
very high. And single page writeback from the LRU is almost
always a bad flush decision.
To complicate matters, filesystems respond very differently to requests
from reclaim according to Christoph Hellwig;
xfs tries to write it back if the requester is kswapd
ext4 ignores the request if it's a delayed allocation
btrfs ignores the request
As a result, each filesystem has different performance characteristics
when under memory pressure and there are many pages being dirtied. In
some cases, the request is ignored entirely so the VM cannot depend on the
IO being dispatched.
The objective of this series is to reduce writing of filesystem-backed
pages from reclaim, play nicely with writeback that is already in progress
and throttle reclaim appropriately when writeback pages are encountered.
The assumption is that the flushers will always write pages faster than if
reclaim issues the IO.
A secondary goal is to avoid the problem whereby direct reclaim splices
two potentially deep call stacks together.
There is a potential new problem as reclaim has less control over how long
before a page in a particularly zone or container is cleaned and direct
reclaimers depend on kswapd or flusher threads to do the necessary work.
However, as filesystems sometimes ignore direct reclaim requests already,
it is not expected to be a serious issue.
Patch 1 disables writeback of filesystem pages from direct reclaim
entirely. Anonymous pages are still written.
Patch 2 removes dead code in lumpy reclaim as it is no longer able
to synchronously write pages. This hurts lumpy reclaim but
there is an expectation that compaction is used for hugepage
allocations these days and lumpy reclaim's days are numbered.
Patches 3-4 add warnings to XFS and ext4 if called from
direct reclaim. With patch 1, this "never happens" and is
intended to catch regressions in this logic in the future.
Patch 5 disables writeback of filesystem pages from kswapd unless
the priority is raised to the point where kswapd is considered
to be in trouble.
Patch 6 throttles reclaimers if too many dirty pages are being
encountered and the zones or backing devices are congested.
Patch 7 invalidates dirty pages found at the end of the LRU so they
are reclaimed quickly after being written back rather than
waiting for a reclaimer to find them
I consider this series to be orthogonal to the writeback work but it is
worth noting that the writeback work affects the viability of patch 8 in
particular.
I tested this on ext4 and xfs using fs_mark, a simple writeback test based
on dd and a micro benchmark that does a streaming write to a large mapping
(exercises use-once LRU logic) followed by streaming writes to a mix of
anonymous and file-backed mappings. The command line for fs_mark when
botted with 512M looked something like
./fs_mark -d /tmp/fsmark-2676 -D 100 -N 150 -n 150 -L 25 -t 1 -S0 -s 10485760
The number of files was adjusted depending on the amount of available
memory so that the files created was about 3xRAM. For multiple threads,
the -d switch is specified multiple times.
The test machine is x86-64 with an older generation of AMD processor with
4 cores. The underlying storage was 4 disks configured as RAID-0 as this
was the best configuration of storage I had available. Swap is on a
separate disk. Dirty ratio was tuned to 40% instead of the default of
20%.
Testing was run with and without monitors to both verify that the patches
were operating as expected and that any performance gain was real and not
due to interference from monitors.
Here is a summary of results based on testing XFS.
512M1P-xfs Files/s mean 32.69 ( 0.00%) 34.44 ( 5.08%)
512M1P-xfs Elapsed Time fsmark 51.41 48.29
512M1P-xfs Elapsed Time simple-wb 114.09 108.61
512M1P-xfs Elapsed Time mmap-strm 113.46 109.34
512M1P-xfs Kswapd efficiency fsmark 62% 63%
512M1P-xfs Kswapd efficiency simple-wb 56% 61%
512M1P-xfs Kswapd efficiency mmap-strm 44% 42%
512M-xfs Files/s mean 30.78 ( 0.00%) 35.94 (14.36%)
512M-xfs Elapsed Time fsmark 56.08 48.90
512M-xfs Elapsed Time simple-wb 112.22 98.13
512M-xfs Elapsed Time mmap-strm 219.15 196.67
512M-xfs Kswapd efficiency fsmark 54% 56%
512M-xfs Kswapd efficiency simple-wb 54% 55%
512M-xfs Kswapd efficiency mmap-strm 45% 44%
512M-4X-xfs Files/s mean 30.31 ( 0.00%) 33.33 ( 9.06%)
512M-4X-xfs Elapsed Time fsmark 63.26 55.88
512M-4X-xfs Elapsed Time simple-wb 100.90 90.25
512M-4X-xfs Elapsed Time mmap-strm 261.73 255.38
512M-4X-xfs Kswapd efficiency fsmark 49% 50%
512M-4X-xfs Kswapd efficiency simple-wb 54% 56%
512M-4X-xfs Kswapd efficiency mmap-strm 37% 36%
512M-16X-xfs Files/s mean 60.89 ( 0.00%) 65.22 ( 6.64%)
512M-16X-xfs Elapsed Time fsmark 67.47 58.25
512M-16X-xfs Elapsed Time simple-wb 103.22 90.89
512M-16X-xfs Elapsed Time mmap-strm 237.09 198.82
512M-16X-xfs Kswapd efficiency fsmark 45% 46%
512M-16X-xfs Kswapd efficiency simple-wb 53% 55%
512M-16X-xfs Kswapd efficiency mmap-strm 33% 33%
Up until 512-4X, the FSmark improvements were statistically significant.
For the 4X and 16X tests the results were within standard deviations but
just barely. The time to completion for all tests is improved which is an
important result. In general, kswapd efficiency is not affected by
skipping dirty pages.
1024M1P-xfs Files/s mean 39.09 ( 0.00%) 41.15 ( 5.01%)
1024M1P-xfs Elapsed Time fsmark 84.14 80.41
1024M1P-xfs Elapsed Time simple-wb 210.77 184.78
1024M1P-xfs Elapsed Time mmap-strm 162.00 160.34
1024M1P-xfs Kswapd efficiency fsmark 69% 75%
1024M1P-xfs Kswapd efficiency simple-wb 71% 77%
1024M1P-xfs Kswapd efficiency mmap-strm 43% 44%
1024M-xfs Files/s mean 35.45 ( 0.00%) 37.00 ( 4.19%)
1024M-xfs Elapsed Time fsmark 94.59 91.00
1024M-xfs Elapsed Time simple-wb 229.84 195.08
1024M-xfs Elapsed Time mmap-strm 405.38 440.29
1024M-xfs Kswapd efficiency fsmark 79% 71%
1024M-xfs Kswapd efficiency simple-wb 74% 74%
1024M-xfs Kswapd efficiency mmap-strm 39% 42%
1024M-4X-xfs Files/s mean 32.63 ( 0.00%) 35.05 ( 6.90%)
1024M-4X-xfs Elapsed Time fsmark 103.33 97.74
1024M-4X-xfs Elapsed Time simple-wb 204.48 178.57
1024M-4X-xfs Elapsed Time mmap-strm 528.38 511.88
1024M-4X-xfs Kswapd efficiency fsmark 81% 70%
1024M-4X-xfs Kswapd efficiency simple-wb 73% 72%
1024M-4X-xfs Kswapd efficiency mmap-strm 39% 38%
1024M-16X-xfs Files/s mean 42.65 ( 0.00%) 42.97 ( 0.74%)
1024M-16X-xfs Elapsed Time fsmark 103.11 99.11
1024M-16X-xfs Elapsed Time simple-wb 200.83 178.24
1024M-16X-xfs Elapsed Time mmap-strm 397.35 459.82
1024M-16X-xfs Kswapd efficiency fsmark 84% 69%
1024M-16X-xfs Kswapd efficiency simple-wb 74% 73%
1024M-16X-xfs Kswapd efficiency mmap-strm 39% 40%
All FSMark tests up to 16X had statistically significant improvements.
For the most part, tests are completing faster with the exception of the
streaming writes to a mixture of anonymous and file-backed mappings which
were slower in two cases
In the cases where the mmap-strm tests were slower, there was more
swapping due to dirty pages being skipped. The number of additional pages
swapped is almost identical to the fewer number of pages written from
reclaim. In other words, roughly the same number of pages were reclaimed
but swapping was slower. As the test is a bit unrealistic and stresses
memory heavily, the small shift is acceptable.
4608M1P-xfs Files/s mean 29.75 ( 0.00%) 30.96 ( 3.91%)
4608M1P-xfs Elapsed Time fsmark 512.01 492.15
4608M1P-xfs Elapsed Time simple-wb 618.18 566.24
4608M1P-xfs Elapsed Time mmap-strm 488.05 465.07
4608M1P-xfs Kswapd efficiency fsmark 93% 86%
4608M1P-xfs Kswapd efficiency simple-wb 88% 84%
4608M1P-xfs Kswapd efficiency mmap-strm 46% 45%
4608M-xfs Files/s mean 27.60 ( 0.00%) 28.85 ( 4.33%)
4608M-xfs Elapsed Time fsmark 555.96 532.34
4608M-xfs Elapsed Time simple-wb 659.72 571.85
4608M-xfs Elapsed Time mmap-strm 1082.57 1146.38
4608M-xfs Kswapd efficiency fsmark 89% 91%
4608M-xfs Kswapd efficiency simple-wb 88% 82%
4608M-xfs Kswapd efficiency mmap-strm 48% 46%
4608M-4X-xfs Files/s mean 26.00 ( 0.00%) 27.47 ( 5.35%)
4608M-4X-xfs Elapsed Time fsmark 592.91 564.00
4608M-4X-xfs Elapsed Time simple-wb 616.65 575.07
4608M-4X-xfs Elapsed Time mmap-strm 1773.02 1631.53
4608M-4X-xfs Kswapd efficiency fsmark 90% 94%
4608M-4X-xfs Kswapd efficiency simple-wb 87% 82%
4608M-4X-xfs Kswapd efficiency mmap-strm 43% 43%
4608M-16X-xfs Files/s mean 26.07 ( 0.00%) 26.42 ( 1.32%)
4608M-16X-xfs Elapsed Time fsmark 602.69 585.78
4608M-16X-xfs Elapsed Time simple-wb 606.60 573.81
4608M-16X-xfs Elapsed Time mmap-strm 1549.75 1441.86
4608M-16X-xfs Kswapd efficiency fsmark 98% 98%
4608M-16X-xfs Kswapd efficiency simple-wb 88% 82%
4608M-16X-xfs Kswapd efficiency mmap-strm 44% 42%
Unlike the other tests, the fsmark results are not statistically
significant but the min and max times are both improved and for the most
part, tests completed faster.
There are other indications that this is an improvement as well. For
example, in the vast majority of cases, there were fewer pages scanned by
direct reclaim implying in many cases that stalls due to direct reclaim
are reduced. KSwapd is scanning more due to skipping dirty pages which is
unfortunate but the CPU usage is still acceptable
In an earlier set of tests, I used blktrace and in almost all cases
throughput throughout the entire test was higher. However, I ended up
discarding those results as recording blktrace data was too heavy for my
liking.
On a laptop, I plugged in a USB stick and ran a similar tests of tests
using it as backing storage. A desktop environment was running and for
the entire duration of the tests, firefox and gnome terminal were
launching and exiting to vaguely simulate a user.
1024M-xfs Files/s mean 0.41 ( 0.00%) 0.44 ( 6.82%)
1024M-xfs Elapsed Time fsmark 2053.52 1641.03
1024M-xfs Elapsed Time simple-wb 1229.53 768.05
1024M-xfs Elapsed Time mmap-strm 4126.44 4597.03
1024M-xfs Kswapd efficiency fsmark 84% 85%
1024M-xfs Kswapd efficiency simple-wb 92% 81%
1024M-xfs Kswapd efficiency mmap-strm 60% 51%
1024M-xfs Avg wait ms fsmark 5404.53 4473.87
1024M-xfs Avg wait ms simple-wb 2541.35 1453.54
1024M-xfs Avg wait ms mmap-strm 3400.25 3852.53
The mmap-strm results were hurt because firefox launching had a tendency
to push the test out of memory. On the postive side, firefox launched
marginally faster with the patches applied. Time to completion for many
tests was faster but more importantly - the "Avg wait" time as measured by
iostat was far lower implying the system would be more responsive. It was
also the case that "Avg wait ms" on the root filesystem was lower. I
tested it manually and while the system felt slightly more responsive
while copying data to a USB stick, it was marginal enough that it could be
my imagination.
This patch: do not writeback filesystem pages in direct reclaim.
When kswapd is failing to keep zones above the min watermark, a process
will enter direct reclaim in the same manner kswapd does. If a dirty page
is encountered during the scan, this page is written to backing storage
using mapping->writepage.
This causes two problems. First, it can result in very deep call stacks,
particularly if the target storage or filesystem are complex. Some
filesystems ignore write requests from direct reclaim as a result. The
second is that a single-page flush is inefficient in terms of IO. While
there is an expectation that the elevator will merge requests, this does
not always happen. Quoting Christoph Hellwig;
The elevator has a relatively small window it can operate on,
and can never fix up a bad large scale writeback pattern.
This patch prevents direct reclaim writing back filesystem pages by
checking if current is kswapd. Anonymous pages are still written to swap
as there is not the equivalent of a flusher thread for anonymous pages.
If the dirty pages cannot be written back, they are placed back on the LRU
lists. There is now a direct dependency on dirty page balancing to
prevent too many pages in the system being dirtied which would prevent
reclaim making forward progress.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Alex Elder <aelder@sgi.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 00:07:38 +00:00
|
|
|
/*
|
2022-05-13 03:23:02 +00:00
|
|
|
* Only kswapd can writeback filesystem folios
|
2017-02-24 22:56:20 +00:00
|
|
|
* to avoid risk of stack overflow. But avoid
|
2022-05-13 03:23:02 +00:00
|
|
|
* injecting inefficient single-folio I/O into
|
2017-02-24 22:56:20 +00:00
|
|
|
* flusher writeback as much as possible: only
|
2022-05-13 03:23:02 +00:00
|
|
|
* write folios when we've encountered many
|
|
|
|
* dirty folios, and when we've already scanned
|
|
|
|
* the rest of the LRU for clean folios and see
|
|
|
|
* the same dirty folios again (with the reclaim
|
|
|
|
* flag set).
|
mm: vmscan: do not writeback filesystem pages in direct reclaim
Testing from the XFS folk revealed that there is still too much I/O from
the end of the LRU in kswapd. Previously it was considered acceptable by
VM people for a small number of pages to be written back from reclaim with
testing generally showing about 0.3% of pages reclaimed were written back
(higher if memory was low). That writing back a small number of pages is
ok has been heavily disputed for quite some time and Dave Chinner
explained it well;
It doesn't have to be a very high number to be a problem. IO
is orders of magnitude slower than the CPU time it takes to
flush a page, so the cost of making a bad flush decision is
very high. And single page writeback from the LRU is almost
always a bad flush decision.
To complicate matters, filesystems respond very differently to requests
from reclaim according to Christoph Hellwig;
xfs tries to write it back if the requester is kswapd
ext4 ignores the request if it's a delayed allocation
btrfs ignores the request
As a result, each filesystem has different performance characteristics
when under memory pressure and there are many pages being dirtied. In
some cases, the request is ignored entirely so the VM cannot depend on the
IO being dispatched.
The objective of this series is to reduce writing of filesystem-backed
pages from reclaim, play nicely with writeback that is already in progress
and throttle reclaim appropriately when writeback pages are encountered.
The assumption is that the flushers will always write pages faster than if
reclaim issues the IO.
A secondary goal is to avoid the problem whereby direct reclaim splices
two potentially deep call stacks together.
There is a potential new problem as reclaim has less control over how long
before a page in a particularly zone or container is cleaned and direct
reclaimers depend on kswapd or flusher threads to do the necessary work.
However, as filesystems sometimes ignore direct reclaim requests already,
it is not expected to be a serious issue.
Patch 1 disables writeback of filesystem pages from direct reclaim
entirely. Anonymous pages are still written.
Patch 2 removes dead code in lumpy reclaim as it is no longer able
to synchronously write pages. This hurts lumpy reclaim but
there is an expectation that compaction is used for hugepage
allocations these days and lumpy reclaim's days are numbered.
Patches 3-4 add warnings to XFS and ext4 if called from
direct reclaim. With patch 1, this "never happens" and is
intended to catch regressions in this logic in the future.
Patch 5 disables writeback of filesystem pages from kswapd unless
the priority is raised to the point where kswapd is considered
to be in trouble.
Patch 6 throttles reclaimers if too many dirty pages are being
encountered and the zones or backing devices are congested.
Patch 7 invalidates dirty pages found at the end of the LRU so they
are reclaimed quickly after being written back rather than
waiting for a reclaimer to find them
I consider this series to be orthogonal to the writeback work but it is
worth noting that the writeback work affects the viability of patch 8 in
particular.
I tested this on ext4 and xfs using fs_mark, a simple writeback test based
on dd and a micro benchmark that does a streaming write to a large mapping
(exercises use-once LRU logic) followed by streaming writes to a mix of
anonymous and file-backed mappings. The command line for fs_mark when
botted with 512M looked something like
./fs_mark -d /tmp/fsmark-2676 -D 100 -N 150 -n 150 -L 25 -t 1 -S0 -s 10485760
The number of files was adjusted depending on the amount of available
memory so that the files created was about 3xRAM. For multiple threads,
the -d switch is specified multiple times.
The test machine is x86-64 with an older generation of AMD processor with
4 cores. The underlying storage was 4 disks configured as RAID-0 as this
was the best configuration of storage I had available. Swap is on a
separate disk. Dirty ratio was tuned to 40% instead of the default of
20%.
Testing was run with and without monitors to both verify that the patches
were operating as expected and that any performance gain was real and not
due to interference from monitors.
Here is a summary of results based on testing XFS.
512M1P-xfs Files/s mean 32.69 ( 0.00%) 34.44 ( 5.08%)
512M1P-xfs Elapsed Time fsmark 51.41 48.29
512M1P-xfs Elapsed Time simple-wb 114.09 108.61
512M1P-xfs Elapsed Time mmap-strm 113.46 109.34
512M1P-xfs Kswapd efficiency fsmark 62% 63%
512M1P-xfs Kswapd efficiency simple-wb 56% 61%
512M1P-xfs Kswapd efficiency mmap-strm 44% 42%
512M-xfs Files/s mean 30.78 ( 0.00%) 35.94 (14.36%)
512M-xfs Elapsed Time fsmark 56.08 48.90
512M-xfs Elapsed Time simple-wb 112.22 98.13
512M-xfs Elapsed Time mmap-strm 219.15 196.67
512M-xfs Kswapd efficiency fsmark 54% 56%
512M-xfs Kswapd efficiency simple-wb 54% 55%
512M-xfs Kswapd efficiency mmap-strm 45% 44%
512M-4X-xfs Files/s mean 30.31 ( 0.00%) 33.33 ( 9.06%)
512M-4X-xfs Elapsed Time fsmark 63.26 55.88
512M-4X-xfs Elapsed Time simple-wb 100.90 90.25
512M-4X-xfs Elapsed Time mmap-strm 261.73 255.38
512M-4X-xfs Kswapd efficiency fsmark 49% 50%
512M-4X-xfs Kswapd efficiency simple-wb 54% 56%
512M-4X-xfs Kswapd efficiency mmap-strm 37% 36%
512M-16X-xfs Files/s mean 60.89 ( 0.00%) 65.22 ( 6.64%)
512M-16X-xfs Elapsed Time fsmark 67.47 58.25
512M-16X-xfs Elapsed Time simple-wb 103.22 90.89
512M-16X-xfs Elapsed Time mmap-strm 237.09 198.82
512M-16X-xfs Kswapd efficiency fsmark 45% 46%
512M-16X-xfs Kswapd efficiency simple-wb 53% 55%
512M-16X-xfs Kswapd efficiency mmap-strm 33% 33%
Up until 512-4X, the FSmark improvements were statistically significant.
For the 4X and 16X tests the results were within standard deviations but
just barely. The time to completion for all tests is improved which is an
important result. In general, kswapd efficiency is not affected by
skipping dirty pages.
1024M1P-xfs Files/s mean 39.09 ( 0.00%) 41.15 ( 5.01%)
1024M1P-xfs Elapsed Time fsmark 84.14 80.41
1024M1P-xfs Elapsed Time simple-wb 210.77 184.78
1024M1P-xfs Elapsed Time mmap-strm 162.00 160.34
1024M1P-xfs Kswapd efficiency fsmark 69% 75%
1024M1P-xfs Kswapd efficiency simple-wb 71% 77%
1024M1P-xfs Kswapd efficiency mmap-strm 43% 44%
1024M-xfs Files/s mean 35.45 ( 0.00%) 37.00 ( 4.19%)
1024M-xfs Elapsed Time fsmark 94.59 91.00
1024M-xfs Elapsed Time simple-wb 229.84 195.08
1024M-xfs Elapsed Time mmap-strm 405.38 440.29
1024M-xfs Kswapd efficiency fsmark 79% 71%
1024M-xfs Kswapd efficiency simple-wb 74% 74%
1024M-xfs Kswapd efficiency mmap-strm 39% 42%
1024M-4X-xfs Files/s mean 32.63 ( 0.00%) 35.05 ( 6.90%)
1024M-4X-xfs Elapsed Time fsmark 103.33 97.74
1024M-4X-xfs Elapsed Time simple-wb 204.48 178.57
1024M-4X-xfs Elapsed Time mmap-strm 528.38 511.88
1024M-4X-xfs Kswapd efficiency fsmark 81% 70%
1024M-4X-xfs Kswapd efficiency simple-wb 73% 72%
1024M-4X-xfs Kswapd efficiency mmap-strm 39% 38%
1024M-16X-xfs Files/s mean 42.65 ( 0.00%) 42.97 ( 0.74%)
1024M-16X-xfs Elapsed Time fsmark 103.11 99.11
1024M-16X-xfs Elapsed Time simple-wb 200.83 178.24
1024M-16X-xfs Elapsed Time mmap-strm 397.35 459.82
1024M-16X-xfs Kswapd efficiency fsmark 84% 69%
1024M-16X-xfs Kswapd efficiency simple-wb 74% 73%
1024M-16X-xfs Kswapd efficiency mmap-strm 39% 40%
All FSMark tests up to 16X had statistically significant improvements.
For the most part, tests are completing faster with the exception of the
streaming writes to a mixture of anonymous and file-backed mappings which
were slower in two cases
In the cases where the mmap-strm tests were slower, there was more
swapping due to dirty pages being skipped. The number of additional pages
swapped is almost identical to the fewer number of pages written from
reclaim. In other words, roughly the same number of pages were reclaimed
but swapping was slower. As the test is a bit unrealistic and stresses
memory heavily, the small shift is acceptable.
4608M1P-xfs Files/s mean 29.75 ( 0.00%) 30.96 ( 3.91%)
4608M1P-xfs Elapsed Time fsmark 512.01 492.15
4608M1P-xfs Elapsed Time simple-wb 618.18 566.24
4608M1P-xfs Elapsed Time mmap-strm 488.05 465.07
4608M1P-xfs Kswapd efficiency fsmark 93% 86%
4608M1P-xfs Kswapd efficiency simple-wb 88% 84%
4608M1P-xfs Kswapd efficiency mmap-strm 46% 45%
4608M-xfs Files/s mean 27.60 ( 0.00%) 28.85 ( 4.33%)
4608M-xfs Elapsed Time fsmark 555.96 532.34
4608M-xfs Elapsed Time simple-wb 659.72 571.85
4608M-xfs Elapsed Time mmap-strm 1082.57 1146.38
4608M-xfs Kswapd efficiency fsmark 89% 91%
4608M-xfs Kswapd efficiency simple-wb 88% 82%
4608M-xfs Kswapd efficiency mmap-strm 48% 46%
4608M-4X-xfs Files/s mean 26.00 ( 0.00%) 27.47 ( 5.35%)
4608M-4X-xfs Elapsed Time fsmark 592.91 564.00
4608M-4X-xfs Elapsed Time simple-wb 616.65 575.07
4608M-4X-xfs Elapsed Time mmap-strm 1773.02 1631.53
4608M-4X-xfs Kswapd efficiency fsmark 90% 94%
4608M-4X-xfs Kswapd efficiency simple-wb 87% 82%
4608M-4X-xfs Kswapd efficiency mmap-strm 43% 43%
4608M-16X-xfs Files/s mean 26.07 ( 0.00%) 26.42 ( 1.32%)
4608M-16X-xfs Elapsed Time fsmark 602.69 585.78
4608M-16X-xfs Elapsed Time simple-wb 606.60 573.81
4608M-16X-xfs Elapsed Time mmap-strm 1549.75 1441.86
4608M-16X-xfs Kswapd efficiency fsmark 98% 98%
4608M-16X-xfs Kswapd efficiency simple-wb 88% 82%
4608M-16X-xfs Kswapd efficiency mmap-strm 44% 42%
Unlike the other tests, the fsmark results are not statistically
significant but the min and max times are both improved and for the most
part, tests completed faster.
There are other indications that this is an improvement as well. For
example, in the vast majority of cases, there were fewer pages scanned by
direct reclaim implying in many cases that stalls due to direct reclaim
are reduced. KSwapd is scanning more due to skipping dirty pages which is
unfortunate but the CPU usage is still acceptable
In an earlier set of tests, I used blktrace and in almost all cases
throughput throughout the entire test was higher. However, I ended up
discarding those results as recording blktrace data was too heavy for my
liking.
On a laptop, I plugged in a USB stick and ran a similar tests of tests
using it as backing storage. A desktop environment was running and for
the entire duration of the tests, firefox and gnome terminal were
launching and exiting to vaguely simulate a user.
1024M-xfs Files/s mean 0.41 ( 0.00%) 0.44 ( 6.82%)
1024M-xfs Elapsed Time fsmark 2053.52 1641.03
1024M-xfs Elapsed Time simple-wb 1229.53 768.05
1024M-xfs Elapsed Time mmap-strm 4126.44 4597.03
1024M-xfs Kswapd efficiency fsmark 84% 85%
1024M-xfs Kswapd efficiency simple-wb 92% 81%
1024M-xfs Kswapd efficiency mmap-strm 60% 51%
1024M-xfs Avg wait ms fsmark 5404.53 4473.87
1024M-xfs Avg wait ms simple-wb 2541.35 1453.54
1024M-xfs Avg wait ms mmap-strm 3400.25 3852.53
The mmap-strm results were hurt because firefox launching had a tendency
to push the test out of memory. On the postive side, firefox launched
marginally faster with the patches applied. Time to completion for many
tests was faster but more importantly - the "Avg wait" time as measured by
iostat was far lower implying the system would be more responsive. It was
also the case that "Avg wait ms" on the root filesystem was lower. I
tested it manually and while the system felt slightly more responsive
while copying data to a USB stick, it was marginal enough that it could be
my imagination.
This patch: do not writeback filesystem pages in direct reclaim.
When kswapd is failing to keep zones above the min watermark, a process
will enter direct reclaim in the same manner kswapd does. If a dirty page
is encountered during the scan, this page is written to backing storage
using mapping->writepage.
This causes two problems. First, it can result in very deep call stacks,
particularly if the target storage or filesystem are complex. Some
filesystems ignore write requests from direct reclaim as a result. The
second is that a single-page flush is inefficient in terms of IO. While
there is an expectation that the elevator will merge requests, this does
not always happen. Quoting Christoph Hellwig;
The elevator has a relatively small window it can operate on,
and can never fix up a bad large scale writeback pattern.
This patch prevents direct reclaim writing back filesystem pages by
checking if current is kswapd. Anonymous pages are still written to swap
as there is not the equivalent of a flusher thread for anonymous pages.
If the dirty pages cannot be written back, they are placed back on the LRU
lists. There is now a direct dependency on dirty page balancing to
prevent too many pages in the system being dirtied which would prevent
reclaim making forward progress.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Alex Elder <aelder@sgi.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 00:07:38 +00:00
|
|
|
*/
|
2022-05-13 03:23:02 +00:00
|
|
|
if (folio_is_file_lru(folio) &&
|
|
|
|
(!current_is_kswapd() ||
|
|
|
|
!folio_test_reclaim(folio) ||
|
2017-02-24 22:56:20 +00:00
|
|
|
!test_bit(PGDAT_DIRTY, &pgdat->flags))) {
|
2011-11-01 00:07:59 +00:00
|
|
|
/*
|
|
|
|
* Immediately reclaim when written back.
|
2022-12-21 18:08:48 +00:00
|
|
|
* Similar in principle to folio_deactivate()
|
2022-05-13 03:23:02 +00:00
|
|
|
* except we already have the folio isolated
|
2011-11-01 00:07:59 +00:00
|
|
|
* and know it's dirty
|
|
|
|
*/
|
2022-05-13 03:23:02 +00:00
|
|
|
node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
|
|
|
|
nr_pages);
|
|
|
|
folio_set_reclaim(folio);
|
2011-11-01 00:07:59 +00:00
|
|
|
|
mm: vmscan: move dirty pages out of the way until they're flushed
We noticed a performance regression when moving hadoop workloads from
3.10 kernels to 4.0 and 4.6. This is accompanied by increased pageout
activity initiated by kswapd as well as frequent bursts of allocation
stalls and direct reclaim scans. Even lowering the dirty ratios to the
equivalent of less than 1% of memory would not eliminate the issue,
suggesting that dirty pages concentrate where the scanner is looking.
This can be traced back to recent efforts of thrash avoidance. Where
3.10 would not detect refaulting pages and continuously supply clean
cache to the inactive list, a thrashing workload on 4.0+ will detect and
activate refaulting pages right away, distilling used-once pages on the
inactive list much more effectively. This is by design, and it makes
sense for clean cache. But for the most part our workload's cache
faults are refaults and its use-once cache is from streaming writes. We
end up with most of the inactive list dirty, and we don't go after the
active cache as long as we have use-once pages around.
But waiting for writes to avoid reclaiming clean cache that *might*
refault is a bad trade-off. Even if the refaults happen, reads are
faster than writes. Before getting bogged down on writeback, reclaim
should first look at *all* cache in the system, even active cache.
To accomplish this, activate pages that are dirty or under writeback
when they reach the end of the inactive LRU. The pages are marked for
immediate reclaim, meaning they'll get moved back to the inactive LRU
tail as soon as they're written back and become reclaimable. But in the
meantime, by reducing the inactive list to only immediately reclaimable
pages, we allow the scanner to deactivate and refill the inactive list
with clean cache from the active list tail to guarantee forward
progress.
[hannes@cmpxchg.org: update comment]
Link: http://lkml.kernel.org/r/20170202191957.22872-8-hannes@cmpxchg.org
Link: http://lkml.kernel.org/r/20170123181641.23938-6-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-24 22:56:23 +00:00
|
|
|
goto activate_locked;
|
mm: vmscan: do not writeback filesystem pages in direct reclaim
Testing from the XFS folk revealed that there is still too much I/O from
the end of the LRU in kswapd. Previously it was considered acceptable by
VM people for a small number of pages to be written back from reclaim with
testing generally showing about 0.3% of pages reclaimed were written back
(higher if memory was low). That writing back a small number of pages is
ok has been heavily disputed for quite some time and Dave Chinner
explained it well;
It doesn't have to be a very high number to be a problem. IO
is orders of magnitude slower than the CPU time it takes to
flush a page, so the cost of making a bad flush decision is
very high. And single page writeback from the LRU is almost
always a bad flush decision.
To complicate matters, filesystems respond very differently to requests
from reclaim according to Christoph Hellwig;
xfs tries to write it back if the requester is kswapd
ext4 ignores the request if it's a delayed allocation
btrfs ignores the request
As a result, each filesystem has different performance characteristics
when under memory pressure and there are many pages being dirtied. In
some cases, the request is ignored entirely so the VM cannot depend on the
IO being dispatched.
The objective of this series is to reduce writing of filesystem-backed
pages from reclaim, play nicely with writeback that is already in progress
and throttle reclaim appropriately when writeback pages are encountered.
The assumption is that the flushers will always write pages faster than if
reclaim issues the IO.
A secondary goal is to avoid the problem whereby direct reclaim splices
two potentially deep call stacks together.
There is a potential new problem as reclaim has less control over how long
before a page in a particularly zone or container is cleaned and direct
reclaimers depend on kswapd or flusher threads to do the necessary work.
However, as filesystems sometimes ignore direct reclaim requests already,
it is not expected to be a serious issue.
Patch 1 disables writeback of filesystem pages from direct reclaim
entirely. Anonymous pages are still written.
Patch 2 removes dead code in lumpy reclaim as it is no longer able
to synchronously write pages. This hurts lumpy reclaim but
there is an expectation that compaction is used for hugepage
allocations these days and lumpy reclaim's days are numbered.
Patches 3-4 add warnings to XFS and ext4 if called from
direct reclaim. With patch 1, this "never happens" and is
intended to catch regressions in this logic in the future.
Patch 5 disables writeback of filesystem pages from kswapd unless
the priority is raised to the point where kswapd is considered
to be in trouble.
Patch 6 throttles reclaimers if too many dirty pages are being
encountered and the zones or backing devices are congested.
Patch 7 invalidates dirty pages found at the end of the LRU so they
are reclaimed quickly after being written back rather than
waiting for a reclaimer to find them
I consider this series to be orthogonal to the writeback work but it is
worth noting that the writeback work affects the viability of patch 8 in
particular.
I tested this on ext4 and xfs using fs_mark, a simple writeback test based
on dd and a micro benchmark that does a streaming write to a large mapping
(exercises use-once LRU logic) followed by streaming writes to a mix of
anonymous and file-backed mappings. The command line for fs_mark when
botted with 512M looked something like
./fs_mark -d /tmp/fsmark-2676 -D 100 -N 150 -n 150 -L 25 -t 1 -S0 -s 10485760
The number of files was adjusted depending on the amount of available
memory so that the files created was about 3xRAM. For multiple threads,
the -d switch is specified multiple times.
The test machine is x86-64 with an older generation of AMD processor with
4 cores. The underlying storage was 4 disks configured as RAID-0 as this
was the best configuration of storage I had available. Swap is on a
separate disk. Dirty ratio was tuned to 40% instead of the default of
20%.
Testing was run with and without monitors to both verify that the patches
were operating as expected and that any performance gain was real and not
due to interference from monitors.
Here is a summary of results based on testing XFS.
512M1P-xfs Files/s mean 32.69 ( 0.00%) 34.44 ( 5.08%)
512M1P-xfs Elapsed Time fsmark 51.41 48.29
512M1P-xfs Elapsed Time simple-wb 114.09 108.61
512M1P-xfs Elapsed Time mmap-strm 113.46 109.34
512M1P-xfs Kswapd efficiency fsmark 62% 63%
512M1P-xfs Kswapd efficiency simple-wb 56% 61%
512M1P-xfs Kswapd efficiency mmap-strm 44% 42%
512M-xfs Files/s mean 30.78 ( 0.00%) 35.94 (14.36%)
512M-xfs Elapsed Time fsmark 56.08 48.90
512M-xfs Elapsed Time simple-wb 112.22 98.13
512M-xfs Elapsed Time mmap-strm 219.15 196.67
512M-xfs Kswapd efficiency fsmark 54% 56%
512M-xfs Kswapd efficiency simple-wb 54% 55%
512M-xfs Kswapd efficiency mmap-strm 45% 44%
512M-4X-xfs Files/s mean 30.31 ( 0.00%) 33.33 ( 9.06%)
512M-4X-xfs Elapsed Time fsmark 63.26 55.88
512M-4X-xfs Elapsed Time simple-wb 100.90 90.25
512M-4X-xfs Elapsed Time mmap-strm 261.73 255.38
512M-4X-xfs Kswapd efficiency fsmark 49% 50%
512M-4X-xfs Kswapd efficiency simple-wb 54% 56%
512M-4X-xfs Kswapd efficiency mmap-strm 37% 36%
512M-16X-xfs Files/s mean 60.89 ( 0.00%) 65.22 ( 6.64%)
512M-16X-xfs Elapsed Time fsmark 67.47 58.25
512M-16X-xfs Elapsed Time simple-wb 103.22 90.89
512M-16X-xfs Elapsed Time mmap-strm 237.09 198.82
512M-16X-xfs Kswapd efficiency fsmark 45% 46%
512M-16X-xfs Kswapd efficiency simple-wb 53% 55%
512M-16X-xfs Kswapd efficiency mmap-strm 33% 33%
Up until 512-4X, the FSmark improvements were statistically significant.
For the 4X and 16X tests the results were within standard deviations but
just barely. The time to completion for all tests is improved which is an
important result. In general, kswapd efficiency is not affected by
skipping dirty pages.
1024M1P-xfs Files/s mean 39.09 ( 0.00%) 41.15 ( 5.01%)
1024M1P-xfs Elapsed Time fsmark 84.14 80.41
1024M1P-xfs Elapsed Time simple-wb 210.77 184.78
1024M1P-xfs Elapsed Time mmap-strm 162.00 160.34
1024M1P-xfs Kswapd efficiency fsmark 69% 75%
1024M1P-xfs Kswapd efficiency simple-wb 71% 77%
1024M1P-xfs Kswapd efficiency mmap-strm 43% 44%
1024M-xfs Files/s mean 35.45 ( 0.00%) 37.00 ( 4.19%)
1024M-xfs Elapsed Time fsmark 94.59 91.00
1024M-xfs Elapsed Time simple-wb 229.84 195.08
1024M-xfs Elapsed Time mmap-strm 405.38 440.29
1024M-xfs Kswapd efficiency fsmark 79% 71%
1024M-xfs Kswapd efficiency simple-wb 74% 74%
1024M-xfs Kswapd efficiency mmap-strm 39% 42%
1024M-4X-xfs Files/s mean 32.63 ( 0.00%) 35.05 ( 6.90%)
1024M-4X-xfs Elapsed Time fsmark 103.33 97.74
1024M-4X-xfs Elapsed Time simple-wb 204.48 178.57
1024M-4X-xfs Elapsed Time mmap-strm 528.38 511.88
1024M-4X-xfs Kswapd efficiency fsmark 81% 70%
1024M-4X-xfs Kswapd efficiency simple-wb 73% 72%
1024M-4X-xfs Kswapd efficiency mmap-strm 39% 38%
1024M-16X-xfs Files/s mean 42.65 ( 0.00%) 42.97 ( 0.74%)
1024M-16X-xfs Elapsed Time fsmark 103.11 99.11
1024M-16X-xfs Elapsed Time simple-wb 200.83 178.24
1024M-16X-xfs Elapsed Time mmap-strm 397.35 459.82
1024M-16X-xfs Kswapd efficiency fsmark 84% 69%
1024M-16X-xfs Kswapd efficiency simple-wb 74% 73%
1024M-16X-xfs Kswapd efficiency mmap-strm 39% 40%
All FSMark tests up to 16X had statistically significant improvements.
For the most part, tests are completing faster with the exception of the
streaming writes to a mixture of anonymous and file-backed mappings which
were slower in two cases
In the cases where the mmap-strm tests were slower, there was more
swapping due to dirty pages being skipped. The number of additional pages
swapped is almost identical to the fewer number of pages written from
reclaim. In other words, roughly the same number of pages were reclaimed
but swapping was slower. As the test is a bit unrealistic and stresses
memory heavily, the small shift is acceptable.
4608M1P-xfs Files/s mean 29.75 ( 0.00%) 30.96 ( 3.91%)
4608M1P-xfs Elapsed Time fsmark 512.01 492.15
4608M1P-xfs Elapsed Time simple-wb 618.18 566.24
4608M1P-xfs Elapsed Time mmap-strm 488.05 465.07
4608M1P-xfs Kswapd efficiency fsmark 93% 86%
4608M1P-xfs Kswapd efficiency simple-wb 88% 84%
4608M1P-xfs Kswapd efficiency mmap-strm 46% 45%
4608M-xfs Files/s mean 27.60 ( 0.00%) 28.85 ( 4.33%)
4608M-xfs Elapsed Time fsmark 555.96 532.34
4608M-xfs Elapsed Time simple-wb 659.72 571.85
4608M-xfs Elapsed Time mmap-strm 1082.57 1146.38
4608M-xfs Kswapd efficiency fsmark 89% 91%
4608M-xfs Kswapd efficiency simple-wb 88% 82%
4608M-xfs Kswapd efficiency mmap-strm 48% 46%
4608M-4X-xfs Files/s mean 26.00 ( 0.00%) 27.47 ( 5.35%)
4608M-4X-xfs Elapsed Time fsmark 592.91 564.00
4608M-4X-xfs Elapsed Time simple-wb 616.65 575.07
4608M-4X-xfs Elapsed Time mmap-strm 1773.02 1631.53
4608M-4X-xfs Kswapd efficiency fsmark 90% 94%
4608M-4X-xfs Kswapd efficiency simple-wb 87% 82%
4608M-4X-xfs Kswapd efficiency mmap-strm 43% 43%
4608M-16X-xfs Files/s mean 26.07 ( 0.00%) 26.42 ( 1.32%)
4608M-16X-xfs Elapsed Time fsmark 602.69 585.78
4608M-16X-xfs Elapsed Time simple-wb 606.60 573.81
4608M-16X-xfs Elapsed Time mmap-strm 1549.75 1441.86
4608M-16X-xfs Kswapd efficiency fsmark 98% 98%
4608M-16X-xfs Kswapd efficiency simple-wb 88% 82%
4608M-16X-xfs Kswapd efficiency mmap-strm 44% 42%
Unlike the other tests, the fsmark results are not statistically
significant but the min and max times are both improved and for the most
part, tests completed faster.
There are other indications that this is an improvement as well. For
example, in the vast majority of cases, there were fewer pages scanned by
direct reclaim implying in many cases that stalls due to direct reclaim
are reduced. KSwapd is scanning more due to skipping dirty pages which is
unfortunate but the CPU usage is still acceptable
In an earlier set of tests, I used blktrace and in almost all cases
throughput throughout the entire test was higher. However, I ended up
discarding those results as recording blktrace data was too heavy for my
liking.
On a laptop, I plugged in a USB stick and ran a similar tests of tests
using it as backing storage. A desktop environment was running and for
the entire duration of the tests, firefox and gnome terminal were
launching and exiting to vaguely simulate a user.
1024M-xfs Files/s mean 0.41 ( 0.00%) 0.44 ( 6.82%)
1024M-xfs Elapsed Time fsmark 2053.52 1641.03
1024M-xfs Elapsed Time simple-wb 1229.53 768.05
1024M-xfs Elapsed Time mmap-strm 4126.44 4597.03
1024M-xfs Kswapd efficiency fsmark 84% 85%
1024M-xfs Kswapd efficiency simple-wb 92% 81%
1024M-xfs Kswapd efficiency mmap-strm 60% 51%
1024M-xfs Avg wait ms fsmark 5404.53 4473.87
1024M-xfs Avg wait ms simple-wb 2541.35 1453.54
1024M-xfs Avg wait ms mmap-strm 3400.25 3852.53
The mmap-strm results were hurt because firefox launching had a tendency
to push the test out of memory. On the postive side, firefox launched
marginally faster with the patches applied. Time to completion for many
tests was faster but more importantly - the "Avg wait" time as measured by
iostat was far lower implying the system would be more responsive. It was
also the case that "Avg wait ms" on the root filesystem was lower. I
tested it manually and while the system felt slightly more responsive
while copying data to a USB stick, it was marginal enough that it could be
my imagination.
This patch: do not writeback filesystem pages in direct reclaim.
When kswapd is failing to keep zones above the min watermark, a process
will enter direct reclaim in the same manner kswapd does. If a dirty page
is encountered during the scan, this page is written to backing storage
using mapping->writepage.
This causes two problems. First, it can result in very deep call stacks,
particularly if the target storage or filesystem are complex. Some
filesystems ignore write requests from direct reclaim as a result. The
second is that a single-page flush is inefficient in terms of IO. While
there is an expectation that the elevator will merge requests, this does
not always happen. Quoting Christoph Hellwig;
The elevator has a relatively small window it can operate on,
and can never fix up a bad large scale writeback pattern.
This patch prevents direct reclaim writing back filesystem pages by
checking if current is kswapd. Anonymous pages are still written to swap
as there is not the equivalent of a flusher thread for anonymous pages.
If the dirty pages cannot be written back, they are placed back on the LRU
lists. There is now a direct dependency on dirty page balancing to
prevent too many pages in the system being dirtied which would prevent
reclaim making forward progress.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Alex Elder <aelder@sgi.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 00:07:38 +00:00
|
|
|
}
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
if (references == FOLIOREF_RECLAIM_CLEAN)
|
2005-04-16 22:20:36 +00:00
|
|
|
goto keep_locked;
|
2022-05-13 03:23:03 +00:00
|
|
|
if (!may_enter_fs(folio, sc->gfp_mask))
|
2005-04-16 22:20:36 +00:00
|
|
|
goto keep_locked;
|
2006-02-01 11:05:28 +00:00
|
|
|
if (!sc->may_writepage)
|
2005-04-16 22:20:36 +00:00
|
|
|
goto keep_locked;
|
|
|
|
|
2015-09-04 22:47:35 +00:00
|
|
|
/*
|
2022-05-13 03:23:02 +00:00
|
|
|
* Folio is dirty. Flush the TLB if a writable entry
|
|
|
|
* potentially exists to avoid CPU writes after I/O
|
2015-09-04 22:47:35 +00:00
|
|
|
* starts and then write it out here.
|
|
|
|
*/
|
|
|
|
try_to_unmap_flush_dirty();
|
2022-05-10 01:20:49 +00:00
|
|
|
switch (pageout(folio, mapping, &plug)) {
|
2005-04-16 22:20:36 +00:00
|
|
|
case PAGE_KEEP:
|
|
|
|
goto keep_locked;
|
|
|
|
case PAGE_ACTIVATE:
|
|
|
|
goto activate_locked;
|
|
|
|
case PAGE_SUCCESS:
|
2022-01-17 19:46:55 +00:00
|
|
|
stat->nr_pageout += nr_pages;
|
mm: vmscan: reclaim writepage is IO cost
The VM tries to balance reclaim pressure between anon and file so as to
reduce the amount of IO incurred due to the memory shortage. It already
counts refaults and swapins, but in addition it should also count
writepage calls during reclaim.
For swap, this is obvious: it's IO that wouldn't have occurred if the
anonymous memory hadn't been under memory pressure. From a relative
balancing point of view this makes sense as well: even if anon is cold and
reclaimable, a cache that isn't thrashing may have equally cold pages that
don't require IO to reclaim.
For file writeback, it's trickier: some of the reclaim writepage IO would
have likely occurred anyway due to dirty expiration. But not all of it -
premature writeback reduces batching and generates additional writes.
Since the flushers are already woken up by the time the VM starts writing
cache pages one by one, let's assume that we'e likely causing writes that
wouldn't have happened without memory pressure. In addition, the per-page
cost of IO would have probably been much cheaper if written in larger
batches from the flusher thread rather than the single-page-writes from
kswapd.
For our purposes - getting the trend right to accelerate convergence on a
stable state that doesn't require paging at all - this is sufficiently
accurate. If we later wanted to optimize for sustained thrashing, we can
still refine the measurements.
Count all writepage calls from kswapd as IO cost toward the LRU that the
page belongs to.
Why do this dynamically? Don't we know in advance that anon pages require
IO to reclaim, and so could build in a static bias?
First, scanning is not the same as reclaiming. If all the anon pages are
referenced, we may not swap for a while just because we're scanning the
anon list. During this time, however, it's important that we age
anonymous memory and the page cache at the same rate so that their
hot-cold gradients are comparable. Everything else being equal, we still
want to reclaim the coldest memory overall.
Second, we keep copies in swap unless the page changes. If there is
swap-backed data that's mostly read (tmpfs file) and has been swapped out
before, we can reclaim it without incurring additional IO.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@surriel.com>
Link: http://lkml.kernel.org/r/20200520232525.798933-14-hannes@cmpxchg.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:03:09 +00:00
|
|
|
|
2022-05-13 03:23:02 +00:00
|
|
|
if (folio_test_writeback(folio))
|
2012-05-29 22:06:19 +00:00
|
|
|
goto keep;
|
2022-05-13 03:23:02 +00:00
|
|
|
if (folio_test_dirty(folio))
|
2005-04-16 22:20:36 +00:00
|
|
|
goto keep;
|
vmscan: narrow the scenarios in whcih lumpy reclaim uses synchrounous reclaim
shrink_page_list() can decide to give up reclaiming a page under a
number of conditions such as
1. trylock_page() failure
2. page is unevictable
3. zone reclaim and page is mapped
4. PageWriteback() is true
5. page is swapbacked and swap is full
6. add_to_swap() failure
7. page is dirty and gfpmask don't have GFP_IO, GFP_FS
8. page is pinned
9. IO queue is congested
10. pageout() start IO, but not finished
With lumpy reclaim, failures result in entering synchronous lumpy reclaim
but this can be unnecessary. In cases (2), (3), (5), (6), (7) and (8),
there is no point retrying. This patch causes lumpy reclaim to abort when
it is known it will fail.
Case (9) is more interesting. current behavior is,
1. start shrink_page_list(async)
2. found queue_congested()
3. skip pageout write
4. still start shrink_page_list(sync)
5. wait on a lot of pages
6. again, found queue_congested()
7. give up pageout write again
So, it's useless time wasting. However, just skipping page reclaim is
also notgood as x86 allocating a huge page needs 512 pages for example.
It can have more dirty pages than queue congestion threshold (~=128).
After this patch, pageout() behaves as follows;
- If order > PAGE_ALLOC_COSTLY_ORDER
Ignore queue congestion always.
- If order <= PAGE_ALLOC_COSTLY_ORDER
skip write page and disable lumpy reclaim.
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-26 21:21:42 +00:00
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* A synchronous write - probably a ramdisk. Go
|
2022-05-13 03:23:02 +00:00
|
|
|
* ahead and try to reclaim the folio.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2022-05-13 03:23:02 +00:00
|
|
|
if (!folio_trylock(folio))
|
2005-04-16 22:20:36 +00:00
|
|
|
goto keep;
|
2022-05-13 03:23:02 +00:00
|
|
|
if (folio_test_dirty(folio) ||
|
|
|
|
folio_test_writeback(folio))
|
2005-04-16 22:20:36 +00:00
|
|
|
goto keep_locked;
|
2022-05-13 03:23:02 +00:00
|
|
|
mapping = folio_mapping(folio);
|
2020-12-15 03:15:00 +00:00
|
|
|
fallthrough;
|
2005-04-16 22:20:36 +00:00
|
|
|
case PAGE_CLEAN:
|
2022-05-13 03:23:02 +00:00
|
|
|
; /* try to free the folio below */
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2022-05-13 03:23:02 +00:00
|
|
|
* If the folio has buffers, try to free the buffer
|
|
|
|
* mappings associated with this folio. If we succeed
|
|
|
|
* we try to free the folio as well.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2022-05-13 03:23:02 +00:00
|
|
|
* We do this even if the folio is dirty.
|
|
|
|
* filemap_release_folio() does not perform I/O, but it
|
|
|
|
* is possible for a folio to have the dirty flag set,
|
|
|
|
* but it is actually clean (all its buffers are clean).
|
|
|
|
* This happens if the buffers were written out directly,
|
|
|
|
* with submit_bh(). ext3 will do this, as well as
|
|
|
|
* the blockdev mapping. filemap_release_folio() will
|
|
|
|
* discover that cleanness and will drop the buffers
|
|
|
|
* and mark the folio clean - it can be freed.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2022-05-13 03:23:02 +00:00
|
|
|
* Rarely, folios can have buffers and no ->mapping.
|
|
|
|
* These are the folios which were not successfully
|
|
|
|
* invalidated in truncate_cleanup_folio(). We try to
|
|
|
|
* drop those buffers here and if that worked, and the
|
|
|
|
* folio is no longer mapped into process address space
|
|
|
|
* (refcount == 1) it can be freed. Otherwise, leave
|
|
|
|
* the folio on the LRU so it is swappable.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2022-05-13 03:23:02 +00:00
|
|
|
if (folio_has_private(folio)) {
|
|
|
|
if (!filemap_release_folio(folio, sc->gfp_mask))
|
2005-04-16 22:20:36 +00:00
|
|
|
goto activate_locked;
|
2022-05-13 03:23:02 +00:00
|
|
|
if (!mapping && folio_ref_count(folio) == 1) {
|
|
|
|
folio_unlock(folio);
|
|
|
|
if (folio_put_testzero(folio))
|
2008-07-26 02:45:30 +00:00
|
|
|
goto free_it;
|
|
|
|
else {
|
|
|
|
/*
|
|
|
|
* rare race with speculative reference.
|
|
|
|
* the speculative reference will free
|
2022-05-13 03:23:02 +00:00
|
|
|
* this folio shortly, so we may
|
2008-07-26 02:45:30 +00:00
|
|
|
* increment nr_reclaimed here (and
|
|
|
|
* leave it off the LRU).
|
|
|
|
*/
|
2022-05-13 03:23:00 +00:00
|
|
|
nr_reclaimed += nr_pages;
|
2008-07-26 02:45:30 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
|
2017-05-03 21:52:32 +00:00
|
|
|
/* follow __remove_mapping for reference */
|
2022-05-13 03:23:03 +00:00
|
|
|
if (!folio_ref_freeze(folio, 1))
|
2017-05-03 21:52:32 +00:00
|
|
|
goto keep_locked;
|
2021-09-02 21:59:39 +00:00
|
|
|
/*
|
2022-05-13 03:23:03 +00:00
|
|
|
* The folio has only one reference left, which is
|
2021-09-02 21:59:39 +00:00
|
|
|
* from the isolation. After the caller puts the
|
2022-05-13 03:23:03 +00:00
|
|
|
* folio back on the lru and drops the reference, the
|
|
|
|
* folio will be freed anyway. It doesn't matter
|
|
|
|
* which lru it goes on. So we don't bother checking
|
|
|
|
* the dirty flag here.
|
2021-09-02 21:59:39 +00:00
|
|
|
*/
|
2022-05-13 03:23:03 +00:00
|
|
|
count_vm_events(PGLAZYFREED, nr_pages);
|
|
|
|
count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
|
2021-12-23 21:50:23 +00:00
|
|
|
} else if (!mapping || !__remove_mapping(mapping, folio, true,
|
mm: vmscan: detect file thrashing at the reclaim root
We use refault information to determine whether the cache workingset is
stable or transitioning, and dynamically adjust the inactive:active file
LRU ratio so as to maximize protection from one-off cache during stable
periods, and minimize IO during transitions.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, refaults only affect the
local LRU order in the cgroup in which they are occuring. As a result,
cache transitions can take longer in a cgrouped system as the active pages
of sibling cgroups aren't challenged when they should be.
[ Right now, this is somewhat theoretical, because the siblings, under
continued regular reclaim pressure, should eventually run out of
inactive pages - and since inactive:active *size* balancing is also
done on a cgroup-local level, we will challenge the active pages
eventually in most cases. But the next patch will move that relative
size enforcement to the reclaim root as well, and then this patch
here will be necessary to propagate refault pressure to siblings. ]
This patch moves refault detection to the root of reclaim. Instead of
remembering the cgroup owner of an evicted page, remember the cgroup that
caused the reclaim to happen. When refaults later occur, they'll
correctly influence the cross-cgroup LRU order that reclaim follows.
I.e. if global reclaim kicked out pages in some subgroup A/B/C, the
refault of those pages will challenge the global LRU order, and not just
the local order down inside C.
[hannes@cmpxchg.org: use page_memcg() instead of another lookup]
Link: http://lkml.kernel.org/r/20191115160722.GA309754@cmpxchg.org
Link: http://lkml.kernel.org/r/20191107205334.158354-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:59 +00:00
|
|
|
sc->target_mem_cgroup))
|
2017-05-03 21:52:32 +00:00
|
|
|
goto keep_locked;
|
mm: put_and_wait_on_page_locked() while page is migrated
Waiting on a page migration entry has used wait_on_page_locked() all along
since 2006: but you cannot safely wait_on_page_locked() without holding a
reference to the page, and that extra reference is enough to make
migrate_page_move_mapping() fail with -EAGAIN, when a racing task faults
on the entry before migrate_page_move_mapping() gets there.
And that failure is retried nine times, amplifying the pain when trying to
migrate a popular page. With a single persistent faulter, migration
sometimes succeeds; with two or three concurrent faulters, success becomes
much less likely (and the more the page was mapped, the worse the overhead
of unmapping and remapping it on each try).
This is especially a problem for memory offlining, where the outer level
retries forever (or until terminated from userspace), because a heavy
refault workload can trigger an endless loop of migration failures.
wait_on_page_locked() is the wrong tool for the job.
David Herrmann (but was he the first?) noticed this issue in 2014:
https://marc.info/?l=linux-mm&m=140110465608116&w=2
Tim Chen started a thread in August 2017 which appears relevant:
https://marc.info/?l=linux-mm&m=150275941014915&w=2 where Kan Liang went
on to implicate __migration_entry_wait():
https://marc.info/?l=linux-mm&m=150300268411980&w=2 and the thread ended
up with the v4.14 commits: 2554db916586 ("sched/wait: Break up long wake
list walk") 11a19c7b099f ("sched/wait: Introduce wakeup boomark in
wake_up_page_bit")
Baoquan He reported "Memory hotplug softlock issue" 14 November 2018:
https://marc.info/?l=linux-mm&m=154217936431300&w=2
We have all assumed that it is essential to hold a page reference while
waiting on a page lock: partly to guarantee that there is still a struct
page when MEMORY_HOTREMOVE is configured, but also to protect against
reuse of the struct page going to someone who then holds the page locked
indefinitely, when the waiter can reasonably expect timely unlocking.
But in fact, so long as wait_on_page_bit_common() does the put_page(), and
is careful not to rely on struct page contents thereafter, there is no
need to hold a reference to the page while waiting on it. That does mean
that this case cannot go back through the loop: but that's fine for the
page migration case, and even if used more widely, is limited by the "Stop
walking if it's locked" optimization in wake_page_function().
Add interface put_and_wait_on_page_locked() to do this, using "behavior"
enum in place of "lock" arg to wait_on_page_bit_common() to implement it.
No interruptible or killable variant needed yet, but they might follow: I
have a vague notion that reporting -EINTR should take precedence over
return from wait_on_page_bit_common() without knowing the page state, so
arrange it accordingly - but that may be nothing but pedantic.
__migration_entry_wait() still has to take a brief reference to the page,
prior to calling put_and_wait_on_page_locked(): but now that it is dropped
before waiting, the chance of impeding page migration is very much
reduced. Should we perhaps disable preemption across this?
shrink_page_list()'s __ClearPageLocked(): that was a surprise! This
survived a lot of testing before that showed up. PageWaiters may have
been set by wait_on_page_bit_common(), and the reference dropped, just
before shrink_page_list() succeeds in freezing its last page reference: in
such a case, unlock_page() must be used. Follow the suggestion from
Michal Hocko, just revert a978d6f52106 ("mm: unlockless reclaim") now:
that optimization predates PageWaiters, and won't buy much these days; but
we can reinstate it for the !PageWaiters case if anyone notices.
It does raise the question: should vmscan.c's is_page_cache_freeable() and
__remove_mapping() now treat a PageWaiters page as if an extra reference
were held? Perhaps, but I don't think it matters much, since
shrink_page_list() already had to win its trylock_page(), so waiters are
not very common there: I noticed no difference when trying the bigger
change, and it's surely not needed while put_and_wait_on_page_locked() is
only used for page migration.
[willy@infradead.org: add put_and_wait_on_page_locked() kerneldoc]
Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1811261121330.1116@eggly.anvils
Signed-off-by: Hugh Dickins <hughd@google.com>
Reported-by: Baoquan He <bhe@redhat.com>
Tested-by: Baoquan He <bhe@redhat.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Baoquan He <bhe@redhat.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: David Herrmann <dh.herrmann@gmail.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:36:14 +00:00
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
folio_unlock(folio);
|
2008-07-26 02:45:30 +00:00
|
|
|
free_it:
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
/*
|
2022-05-13 03:23:03 +00:00
|
|
|
* Folio may get swapped out as a whole, need to account
|
|
|
|
* all pages in it.
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
*/
|
|
|
|
nr_reclaimed += nr_pages;
|
2010-08-10 00:19:31 +00:00
|
|
|
|
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* Is there need to periodically free_folio_list? It would
|
2010-08-10 00:19:31 +00:00
|
|
|
* appear not as the counts should be low
|
|
|
|
*/
|
2022-05-13 03:23:03 +00:00
|
|
|
if (unlikely(folio_test_large(folio)))
|
2022-06-17 17:50:17 +00:00
|
|
|
destroy_large_folio(folio);
|
2019-09-23 22:38:09 +00:00
|
|
|
else
|
2022-09-02 19:45:57 +00:00
|
|
|
list_add(&folio->lru, &free_folios);
|
2005-04-16 22:20:36 +00:00
|
|
|
continue;
|
|
|
|
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
activate_locked_split:
|
|
|
|
/*
|
|
|
|
* The tail pages that are failed to add into swap cache
|
|
|
|
* reach here. Fixup nr_scanned and nr_pages.
|
|
|
|
*/
|
|
|
|
if (nr_pages > 1) {
|
|
|
|
sc->nr_scanned -= (nr_pages - 1);
|
|
|
|
nr_pages = 1;
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
activate_locked:
|
2008-10-19 03:26:23 +00:00
|
|
|
/* Not a candidate for swapping, so reclaim swap space. */
|
2022-05-13 03:23:03 +00:00
|
|
|
if (folio_test_swapcache(folio) &&
|
2022-09-02 19:46:43 +00:00
|
|
|
(mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
|
2022-09-02 19:46:06 +00:00
|
|
|
folio_free_swap(folio);
|
2022-05-13 03:23:03 +00:00
|
|
|
VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
|
|
|
|
if (!folio_test_mlocked(folio)) {
|
|
|
|
int type = folio_is_file_lru(folio);
|
|
|
|
folio_set_active(folio);
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
stat->nr_activate[type] += nr_pages;
|
2022-05-13 03:23:03 +00:00
|
|
|
count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
|
2017-05-03 21:54:13 +00:00
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
keep_locked:
|
2022-05-13 03:23:03 +00:00
|
|
|
folio_unlock(folio);
|
2005-04-16 22:20:36 +00:00
|
|
|
keep:
|
2022-09-02 19:45:57 +00:00
|
|
|
list_add(&folio->lru, &ret_folios);
|
2022-05-13 03:23:03 +00:00
|
|
|
VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
|
|
|
|
folio_test_unevictable(folio), folio);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2022-09-02 19:45:57 +00:00
|
|
|
/* 'folio_list' is always empty here */
|
2021-09-02 21:59:16 +00:00
|
|
|
|
2022-05-13 03:23:03 +00:00
|
|
|
/* Migrate folios selected for demotion */
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
|
|
|
|
/* Folios that could not be demoted are still in @demote_folios */
|
|
|
|
if (!list_empty(&demote_folios)) {
|
2022-12-01 23:33:17 +00:00
|
|
|
/* Folios which weren't demoted go back on @folio_list */
|
2022-09-02 19:45:57 +00:00
|
|
|
list_splice_init(&demote_folios, folio_list);
|
2022-12-01 23:33:17 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* goto retry to reclaim the undemoted folios in folio_list if
|
|
|
|
* desired.
|
|
|
|
*
|
|
|
|
* Reclaiming directly from top tier nodes is not often desired
|
|
|
|
* due to it breaking the LRU ordering: in general memory
|
|
|
|
* should be reclaimed from lower tier nodes and demoted from
|
|
|
|
* top tier nodes.
|
|
|
|
*
|
|
|
|
* However, disabling reclaim from top tier nodes entirely
|
|
|
|
* would cause ooms in edge scenarios where lower tier memory
|
|
|
|
* is unreclaimable for whatever reason, eg memory being
|
|
|
|
* mlocked or too hot to reclaim. We can disable reclaim
|
|
|
|
* from top tier nodes in proactive reclaim though as that is
|
|
|
|
* not real memory pressure.
|
|
|
|
*/
|
|
|
|
if (!sc->proactive) {
|
|
|
|
do_demote_pass = false;
|
|
|
|
goto retry;
|
|
|
|
}
|
2021-09-02 21:59:16 +00:00
|
|
|
}
|
2010-08-10 00:19:31 +00:00
|
|
|
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
mem_cgroup_uncharge_list(&free_folios);
|
2015-09-04 22:47:32 +00:00
|
|
|
try_to_unmap_flush();
|
2022-09-02 19:45:57 +00:00
|
|
|
free_unref_page_list(&free_folios);
|
2010-08-10 00:19:31 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
list_splice(&ret_folios, folio_list);
|
2019-05-14 00:16:51 +00:00
|
|
|
count_vm_events(PGACTIVATE, pgactivate);
|
2019-03-05 23:48:15 +00:00
|
|
|
|
2022-05-10 01:20:49 +00:00
|
|
|
if (plug)
|
|
|
|
swap_write_unplug(plug);
|
2006-03-22 08:08:20 +00:00
|
|
|
return nr_reclaimed;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2020-06-03 23:01:18 +00:00
|
|
|
unsigned int reclaim_clean_pages_from_list(struct zone *zone,
|
2022-09-02 19:45:57 +00:00
|
|
|
struct list_head *folio_list)
|
2012-10-08 23:31:55 +00:00
|
|
|
{
|
|
|
|
struct scan_control sc = {
|
|
|
|
.gfp_mask = GFP_KERNEL,
|
|
|
|
.may_unmap = 1,
|
|
|
|
};
|
mm/vmscan: count layzfree pages and fix nr_isolated_* mismatch
Fix an nr_isolate_* mismatch problem between cma and dirty lazyfree pages.
If try_to_unmap_one is used for reclaim and it detects a dirty lazyfree
page, then the lazyfree page is changed to a normal anon page having
SwapBacked by commit 802a3a92ad7a ("mm: reclaim MADV_FREE pages"). Even
with the change, reclaim context correctly counts isolated files because
it uses is_file_lru to distinguish file. And the change to anon is not
happened if try_to_unmap_one is used for migration. So migration context
like compaction also correctly counts isolated files even though it uses
page_is_file_lru insted of is_file_lru. Recently page_is_file_cache was
renamed to page_is_file_lru by commit 9de4f22a60f7 ("mm: code cleanup for
MADV_FREE").
But the nr_isolate_* mismatch problem happens on cma alloc. There is
reclaim_clean_pages_from_list which is being used only by cma. It was
introduced by commit 02c6de8d757c ("mm: cma: discard clean pages during
contiguous allocation instead of migration") to reclaim clean file pages
without migration. The cma alloc uses both reclaim_clean_pages_from_list
and migrate_pages, and it uses page_is_file_lru to count isolated files.
If there are dirty lazyfree pages allocated from cma memory region, the
pages are counted as isolated file at the beginging but are counted as
isolated anon after finished.
Mem-Info:
Node 0 active_anon:3045904kB inactive_anon:611448kB active_file:14892kB inactive_file:205636kB unevictable:10416kB isolated(anon):0kB isolated(file):37664kB mapped:630216kB dirty:384kB writeback:0kB shmem:42576kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Like log above, there were too much isolated files, 37664kB, which
triggers too_many_isolated in reclaim even when there is no actually
isolated file in system wide. It could be reproducible by running two
programs, writing on MADV_FREE page and doing cma alloc, respectively.
Although isolated anon is 0, I found that the internal value of isolated
anon was the negative value of isolated file.
Fix this by compensating the isolated count for both LRU lists. Count
non-discarded lazyfree pages in shrink_page_list, then compensate the
counted number in reclaim_clean_pages_from_list.
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200426011718.30246-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:01:15 +00:00
|
|
|
struct reclaim_stat stat;
|
2020-06-03 23:01:18 +00:00
|
|
|
unsigned int nr_reclaimed;
|
2022-06-17 15:42:44 +00:00
|
|
|
struct folio *folio, *next;
|
|
|
|
LIST_HEAD(clean_folios);
|
2021-07-01 01:49:48 +00:00
|
|
|
unsigned int noreclaim_flag;
|
2012-10-08 23:31:55 +00:00
|
|
|
|
2022-06-17 15:42:44 +00:00
|
|
|
list_for_each_entry_safe(folio, next, folio_list, lru) {
|
|
|
|
if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
|
|
|
|
!folio_test_dirty(folio) && !__folio_test_movable(folio) &&
|
|
|
|
!folio_test_unevictable(folio)) {
|
|
|
|
folio_clear_active(folio);
|
|
|
|
list_move(&folio->lru, &clean_folios);
|
2012-10-08 23:31:55 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-07-01 01:49:48 +00:00
|
|
|
/*
|
|
|
|
* We should be safe here since we are only dealing with file pages and
|
|
|
|
* we are not kswapd and therefore cannot write dirty file pages. But
|
|
|
|
* call memalloc_noreclaim_save() anyway, just in case these conditions
|
|
|
|
* change in the future.
|
|
|
|
*/
|
|
|
|
noreclaim_flag = memalloc_noreclaim_save();
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
|
2020-12-15 03:06:39 +00:00
|
|
|
&stat, true);
|
2021-07-01 01:49:48 +00:00
|
|
|
memalloc_noreclaim_restore(noreclaim_flag);
|
|
|
|
|
2022-06-17 15:42:44 +00:00
|
|
|
list_splice(&clean_folios, folio_list);
|
2020-11-14 06:51:46 +00:00
|
|
|
mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
|
|
|
|
-(long)nr_reclaimed);
|
mm/vmscan: count layzfree pages and fix nr_isolated_* mismatch
Fix an nr_isolate_* mismatch problem between cma and dirty lazyfree pages.
If try_to_unmap_one is used for reclaim and it detects a dirty lazyfree
page, then the lazyfree page is changed to a normal anon page having
SwapBacked by commit 802a3a92ad7a ("mm: reclaim MADV_FREE pages"). Even
with the change, reclaim context correctly counts isolated files because
it uses is_file_lru to distinguish file. And the change to anon is not
happened if try_to_unmap_one is used for migration. So migration context
like compaction also correctly counts isolated files even though it uses
page_is_file_lru insted of is_file_lru. Recently page_is_file_cache was
renamed to page_is_file_lru by commit 9de4f22a60f7 ("mm: code cleanup for
MADV_FREE").
But the nr_isolate_* mismatch problem happens on cma alloc. There is
reclaim_clean_pages_from_list which is being used only by cma. It was
introduced by commit 02c6de8d757c ("mm: cma: discard clean pages during
contiguous allocation instead of migration") to reclaim clean file pages
without migration. The cma alloc uses both reclaim_clean_pages_from_list
and migrate_pages, and it uses page_is_file_lru to count isolated files.
If there are dirty lazyfree pages allocated from cma memory region, the
pages are counted as isolated file at the beginging but are counted as
isolated anon after finished.
Mem-Info:
Node 0 active_anon:3045904kB inactive_anon:611448kB active_file:14892kB inactive_file:205636kB unevictable:10416kB isolated(anon):0kB isolated(file):37664kB mapped:630216kB dirty:384kB writeback:0kB shmem:42576kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Like log above, there were too much isolated files, 37664kB, which
triggers too_many_isolated in reclaim even when there is no actually
isolated file in system wide. It could be reproducible by running two
programs, writing on MADV_FREE page and doing cma alloc, respectively.
Although isolated anon is 0, I found that the internal value of isolated
anon was the negative value of isolated file.
Fix this by compensating the isolated count for both LRU lists. Count
non-discarded lazyfree pages in shrink_page_list, then compensate the
counted number in reclaim_clean_pages_from_list.
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200426011718.30246-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:01:15 +00:00
|
|
|
/*
|
|
|
|
* Since lazyfree pages are isolated from file LRU from the beginning,
|
|
|
|
* they will rotate back to anonymous LRU in the end if it failed to
|
|
|
|
* discard so isolated count will be mismatched.
|
|
|
|
* Compensate the isolated count for both LRU lists.
|
|
|
|
*/
|
|
|
|
mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
|
|
|
|
stat.nr_lazyfree_fail);
|
|
|
|
mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
|
2020-11-14 06:51:46 +00:00
|
|
|
-(long)stat.nr_lazyfree_fail);
|
mm/vmscan: count layzfree pages and fix nr_isolated_* mismatch
Fix an nr_isolate_* mismatch problem between cma and dirty lazyfree pages.
If try_to_unmap_one is used for reclaim and it detects a dirty lazyfree
page, then the lazyfree page is changed to a normal anon page having
SwapBacked by commit 802a3a92ad7a ("mm: reclaim MADV_FREE pages"). Even
with the change, reclaim context correctly counts isolated files because
it uses is_file_lru to distinguish file. And the change to anon is not
happened if try_to_unmap_one is used for migration. So migration context
like compaction also correctly counts isolated files even though it uses
page_is_file_lru insted of is_file_lru. Recently page_is_file_cache was
renamed to page_is_file_lru by commit 9de4f22a60f7 ("mm: code cleanup for
MADV_FREE").
But the nr_isolate_* mismatch problem happens on cma alloc. There is
reclaim_clean_pages_from_list which is being used only by cma. It was
introduced by commit 02c6de8d757c ("mm: cma: discard clean pages during
contiguous allocation instead of migration") to reclaim clean file pages
without migration. The cma alloc uses both reclaim_clean_pages_from_list
and migrate_pages, and it uses page_is_file_lru to count isolated files.
If there are dirty lazyfree pages allocated from cma memory region, the
pages are counted as isolated file at the beginging but are counted as
isolated anon after finished.
Mem-Info:
Node 0 active_anon:3045904kB inactive_anon:611448kB active_file:14892kB inactive_file:205636kB unevictable:10416kB isolated(anon):0kB isolated(file):37664kB mapped:630216kB dirty:384kB writeback:0kB shmem:42576kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
Like log above, there were too much isolated files, 37664kB, which
triggers too_many_isolated in reclaim even when there is no actually
isolated file in system wide. It could be reproducible by running two
programs, writing on MADV_FREE page and doing cma alloc, respectively.
Although isolated anon is 0, I found that the internal value of isolated
anon was the negative value of isolated file.
Fix this by compensating the isolated count for both LRU lists. Count
non-discarded lazyfree pages in shrink_page_list, then compensate the
counted number in reclaim_clean_pages_from_list.
Reported-by: Yong-Taek Lee <ytk.lee@samsung.com>
Suggested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200426011718.30246-1-jaewon31.kim@samsung.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:01:15 +00:00
|
|
|
return nr_reclaimed;
|
2012-10-08 23:31:55 +00:00
|
|
|
}
|
|
|
|
|
2016-07-28 22:47:17 +00:00
|
|
|
/*
|
|
|
|
* Update LRU sizes after isolating pages. The LRU size updates must
|
2020-06-04 23:49:10 +00:00
|
|
|
* be complete before mem_cgroup_update_lru_size due to a sanity check.
|
2016-07-28 22:47:17 +00:00
|
|
|
*/
|
|
|
|
static __always_inline void update_lru_sizes(struct lruvec *lruvec,
|
2017-01-11 00:58:04 +00:00
|
|
|
enum lru_list lru, unsigned long *nr_zone_taken)
|
2016-07-28 22:47:17 +00:00
|
|
|
{
|
|
|
|
int zid;
|
|
|
|
|
|
|
|
for (zid = 0; zid < MAX_NR_ZONES; zid++) {
|
|
|
|
if (!nr_zone_taken[zid])
|
|
|
|
continue;
|
|
|
|
|
2020-06-03 23:01:12 +00:00
|
|
|
update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
|
2017-01-11 00:58:04 +00:00
|
|
|
}
|
|
|
|
|
2016-07-28 22:47:17 +00:00
|
|
|
}
|
|
|
|
|
2021-07-01 01:53:19 +00:00
|
|
|
/*
|
2020-12-15 22:21:31 +00:00
|
|
|
* Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
|
|
|
|
*
|
|
|
|
* lruvec->lru_lock is heavily contended. Some of the functions that
|
2005-04-16 22:20:36 +00:00
|
|
|
* shrink the lists perform better by taking out a batch of pages
|
|
|
|
* and working on them outside the LRU lock.
|
|
|
|
*
|
|
|
|
* For pagecache intensive workloads, this function is the hottest
|
|
|
|
* spot in the kernel (apart from copy_*_user functions).
|
|
|
|
*
|
2020-12-15 22:21:31 +00:00
|
|
|
* Lru_lock must be held before calling this function.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
mm: vmscan: scan until it finds eligible pages
Although there are a ton of free swap and anonymous LRU page in elgible
zones, OOM happened.
balloon invoked oom-killer: gfp_mask=0x17080c0(GFP_KERNEL_ACCOUNT|__GFP_ZERO|__GFP_NOTRACK), nodemask=(null), order=0, oom_score_adj=0
CPU: 7 PID: 1138 Comm: balloon Not tainted 4.11.0-rc6-mm1-zram-00289-ge228d67e9677-dirty #17
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
oom_kill_process+0x21d/0x3f0
out_of_memory+0xd8/0x390
__alloc_pages_slowpath+0xbc1/0xc50
__alloc_pages_nodemask+0x1a5/0x1c0
pte_alloc_one+0x20/0x50
__pte_alloc+0x1e/0x110
__handle_mm_fault+0x919/0x960
handle_mm_fault+0x77/0x120
__do_page_fault+0x27a/0x550
trace_do_page_fault+0x43/0x150
do_async_page_fault+0x2c/0x90
async_page_fault+0x28/0x30
Mem-Info:
active_anon:424716 inactive_anon:65314 isolated_anon:0
active_file:52 inactive_file:46 isolated_file:0
unevictable:0 dirty:27 writeback:0 unstable:0
slab_reclaimable:3967 slab_unreclaimable:4125
mapped:133 shmem:43 pagetables:1674 bounce:0
free:4637 free_pcp:225 free_cma:0
Node 0 active_anon:1698864kB inactive_anon:261256kB active_file:208kB inactive_file:184kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:532kB dirty:108kB writeback:0kB shmem:172kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
DMA free:7316kB min:32kB low:44kB high:56kB active_anon:8064kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:15992kB managed:15908kB mlocked:0kB slab_reclaimable:464kB slab_unreclaimable:40kB kernel_stack:0kB pagetables:24kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
lowmem_reserve[]: 0 992 992 1952
DMA32 free:9088kB min:2048kB low:3064kB high:4080kB active_anon:952176kB inactive_anon:0kB active_file:36kB inactive_file:0kB unevictable:0kB writepending:88kB present:1032192kB managed:1019388kB mlocked:0kB slab_reclaimable:13532kB slab_unreclaimable:16460kB kernel_stack:3552kB pagetables:6672kB bounce:0kB free_pcp:56kB local_pcp:24kB free_cma:0kB
lowmem_reserve[]: 0 0 0 959
Movable free:3644kB min:1980kB low:2960kB high:3940kB active_anon:738560kB inactive_anon:261340kB active_file:188kB inactive_file:640kB unevictable:0kB writepending:20kB present:1048444kB managed:1010816kB mlocked:0kB slab_reclaimable:0kB slab_unreclaimable:0kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:832kB local_pcp:60kB free_cma:0kB
lowmem_reserve[]: 0 0 0 0
DMA: 1*4kB (E) 0*8kB 18*16kB (E) 10*32kB (E) 10*64kB (E) 9*128kB (ME) 8*256kB (E) 2*512kB (E) 2*1024kB (E) 0*2048kB 0*4096kB = 7524kB
DMA32: 417*4kB (UMEH) 181*8kB (UMEH) 68*16kB (UMEH) 48*32kB (UMEH) 14*64kB (MH) 3*128kB (M) 1*256kB (H) 1*512kB (M) 2*1024kB (M) 0*2048kB 0*4096kB = 9836kB
Movable: 1*4kB (M) 1*8kB (M) 1*16kB (M) 1*32kB (M) 0*64kB 1*128kB (M) 2*256kB (M) 4*512kB (M) 1*1024kB (M) 0*2048kB 0*4096kB = 3772kB
378 total pagecache pages
17 pages in swap cache
Swap cache stats: add 17325, delete 17302, find 0/27
Free swap = 978940kB
Total swap = 1048572kB
524157 pages RAM
0 pages HighMem/MovableOnly
12629 pages reserved
0 pages cma reserved
0 pages hwpoisoned
[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name
[ 433] 0 433 4904 5 14 3 82 0 upstart-udev-br
[ 438] 0 438 12371 5 27 3 191 -1000 systemd-udevd
With investigation, skipping page of isolate_lru_pages makes reclaim
void because it returns zero nr_taken easily so LRU shrinking is
effectively nothing and just increases priority aggressively. Finally,
OOM happens.
The problem is that get_scan_count determines nr_to_scan with eligible
zones so although priority drops to zero, it couldn't reclaim any pages
if the LRU contains mostly ineligible pages.
get_scan_count:
size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
size = size >> sc->priority;
Assumes sc->priority is 0 and LRU list is as follows.
N-N-N-N-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H
(Ie, small eligible pages are in the head of LRU but others are
almost ineligible pages)
In that case, size becomes 4 so VM want to scan 4 pages but 4 pages from
tail of the LRU are not eligible pages. If get_scan_count counts
skipped pages, it doesn't reclaim any pages remained after scanning 4
pages so it ends up OOM happening.
This patch makes isolate_lru_pages try to scan pages until it encounters
eligible zones's pages.
[akpm@linux-foundation.org: clean up mind-bending `for' statement. Tweak comment text]
Fixes: 3db65812d688 ("Revert "mm, vmscan: account for skipped pages as a partial scan"")
Link: http://lkml.kernel.org/r/1494457232-27401-1-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-12 22:47:06 +00:00
|
|
|
* @nr_to_scan: The number of eligible pages to look through on the list.
|
2012-05-29 22:06:58 +00:00
|
|
|
* @lruvec: The LRU vector to pull pages from.
|
2005-04-16 22:20:36 +00:00
|
|
|
* @dst: The temp list to put pages on to.
|
2012-01-13 01:20:06 +00:00
|
|
|
* @nr_scanned: The number of pages that were scanned.
|
2012-03-21 23:33:51 +00:00
|
|
|
* @sc: The scan_control struct for this reclaim session
|
2012-05-29 22:06:53 +00:00
|
|
|
* @lru: LRU list id for isolating
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* returns how many pages were moved onto *@dst.
|
|
|
|
*/
|
2022-09-02 19:45:57 +00:00
|
|
|
static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
|
2012-05-29 22:06:58 +00:00
|
|
|
struct lruvec *lruvec, struct list_head *dst,
|
2012-03-21 23:33:51 +00:00
|
|
|
unsigned long *nr_scanned, struct scan_control *sc,
|
2019-03-05 23:46:55 +00:00
|
|
|
enum lru_list lru)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2012-05-29 22:07:09 +00:00
|
|
|
struct list_head *src = &lruvec->lists[lru];
|
2006-03-22 08:08:19 +00:00
|
|
|
unsigned long nr_taken = 0;
|
2016-07-28 22:45:31 +00:00
|
|
|
unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
|
2016-07-28 22:46:59 +00:00
|
|
|
unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
|
2017-05-03 21:52:13 +00:00
|
|
|
unsigned long skipped = 0;
|
mm: vmscan: scan until it finds eligible pages
Although there are a ton of free swap and anonymous LRU page in elgible
zones, OOM happened.
balloon invoked oom-killer: gfp_mask=0x17080c0(GFP_KERNEL_ACCOUNT|__GFP_ZERO|__GFP_NOTRACK), nodemask=(null), order=0, oom_score_adj=0
CPU: 7 PID: 1138 Comm: balloon Not tainted 4.11.0-rc6-mm1-zram-00289-ge228d67e9677-dirty #17
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
oom_kill_process+0x21d/0x3f0
out_of_memory+0xd8/0x390
__alloc_pages_slowpath+0xbc1/0xc50
__alloc_pages_nodemask+0x1a5/0x1c0
pte_alloc_one+0x20/0x50
__pte_alloc+0x1e/0x110
__handle_mm_fault+0x919/0x960
handle_mm_fault+0x77/0x120
__do_page_fault+0x27a/0x550
trace_do_page_fault+0x43/0x150
do_async_page_fault+0x2c/0x90
async_page_fault+0x28/0x30
Mem-Info:
active_anon:424716 inactive_anon:65314 isolated_anon:0
active_file:52 inactive_file:46 isolated_file:0
unevictable:0 dirty:27 writeback:0 unstable:0
slab_reclaimable:3967 slab_unreclaimable:4125
mapped:133 shmem:43 pagetables:1674 bounce:0
free:4637 free_pcp:225 free_cma:0
Node 0 active_anon:1698864kB inactive_anon:261256kB active_file:208kB inactive_file:184kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:532kB dirty:108kB writeback:0kB shmem:172kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
DMA free:7316kB min:32kB low:44kB high:56kB active_anon:8064kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:15992kB managed:15908kB mlocked:0kB slab_reclaimable:464kB slab_unreclaimable:40kB kernel_stack:0kB pagetables:24kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
lowmem_reserve[]: 0 992 992 1952
DMA32 free:9088kB min:2048kB low:3064kB high:4080kB active_anon:952176kB inactive_anon:0kB active_file:36kB inactive_file:0kB unevictable:0kB writepending:88kB present:1032192kB managed:1019388kB mlocked:0kB slab_reclaimable:13532kB slab_unreclaimable:16460kB kernel_stack:3552kB pagetables:6672kB bounce:0kB free_pcp:56kB local_pcp:24kB free_cma:0kB
lowmem_reserve[]: 0 0 0 959
Movable free:3644kB min:1980kB low:2960kB high:3940kB active_anon:738560kB inactive_anon:261340kB active_file:188kB inactive_file:640kB unevictable:0kB writepending:20kB present:1048444kB managed:1010816kB mlocked:0kB slab_reclaimable:0kB slab_unreclaimable:0kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:832kB local_pcp:60kB free_cma:0kB
lowmem_reserve[]: 0 0 0 0
DMA: 1*4kB (E) 0*8kB 18*16kB (E) 10*32kB (E) 10*64kB (E) 9*128kB (ME) 8*256kB (E) 2*512kB (E) 2*1024kB (E) 0*2048kB 0*4096kB = 7524kB
DMA32: 417*4kB (UMEH) 181*8kB (UMEH) 68*16kB (UMEH) 48*32kB (UMEH) 14*64kB (MH) 3*128kB (M) 1*256kB (H) 1*512kB (M) 2*1024kB (M) 0*2048kB 0*4096kB = 9836kB
Movable: 1*4kB (M) 1*8kB (M) 1*16kB (M) 1*32kB (M) 0*64kB 1*128kB (M) 2*256kB (M) 4*512kB (M) 1*1024kB (M) 0*2048kB 0*4096kB = 3772kB
378 total pagecache pages
17 pages in swap cache
Swap cache stats: add 17325, delete 17302, find 0/27
Free swap = 978940kB
Total swap = 1048572kB
524157 pages RAM
0 pages HighMem/MovableOnly
12629 pages reserved
0 pages cma reserved
0 pages hwpoisoned
[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name
[ 433] 0 433 4904 5 14 3 82 0 upstart-udev-br
[ 438] 0 438 12371 5 27 3 191 -1000 systemd-udevd
With investigation, skipping page of isolate_lru_pages makes reclaim
void because it returns zero nr_taken easily so LRU shrinking is
effectively nothing and just increases priority aggressively. Finally,
OOM happens.
The problem is that get_scan_count determines nr_to_scan with eligible
zones so although priority drops to zero, it couldn't reclaim any pages
if the LRU contains mostly ineligible pages.
get_scan_count:
size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
size = size >> sc->priority;
Assumes sc->priority is 0 and LRU list is as follows.
N-N-N-N-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H
(Ie, small eligible pages are in the head of LRU but others are
almost ineligible pages)
In that case, size becomes 4 so VM want to scan 4 pages but 4 pages from
tail of the LRU are not eligible pages. If get_scan_count counts
skipped pages, it doesn't reclaim any pages remained after scanning 4
pages so it ends up OOM happening.
This patch makes isolate_lru_pages try to scan pages until it encounters
eligible zones's pages.
[akpm@linux-foundation.org: clean up mind-bending `for' statement. Tweak comment text]
Fixes: 3db65812d688 ("Revert "mm, vmscan: account for skipped pages as a partial scan"")
Link: http://lkml.kernel.org/r/1494457232-27401-1-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-12 22:47:06 +00:00
|
|
|
unsigned long scan, total_scan, nr_pages;
|
2022-06-17 15:42:45 +00:00
|
|
|
LIST_HEAD(folios_skipped);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
total_scan = 0;
|
mm: vmscan: scan until it finds eligible pages
Although there are a ton of free swap and anonymous LRU page in elgible
zones, OOM happened.
balloon invoked oom-killer: gfp_mask=0x17080c0(GFP_KERNEL_ACCOUNT|__GFP_ZERO|__GFP_NOTRACK), nodemask=(null), order=0, oom_score_adj=0
CPU: 7 PID: 1138 Comm: balloon Not tainted 4.11.0-rc6-mm1-zram-00289-ge228d67e9677-dirty #17
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
oom_kill_process+0x21d/0x3f0
out_of_memory+0xd8/0x390
__alloc_pages_slowpath+0xbc1/0xc50
__alloc_pages_nodemask+0x1a5/0x1c0
pte_alloc_one+0x20/0x50
__pte_alloc+0x1e/0x110
__handle_mm_fault+0x919/0x960
handle_mm_fault+0x77/0x120
__do_page_fault+0x27a/0x550
trace_do_page_fault+0x43/0x150
do_async_page_fault+0x2c/0x90
async_page_fault+0x28/0x30
Mem-Info:
active_anon:424716 inactive_anon:65314 isolated_anon:0
active_file:52 inactive_file:46 isolated_file:0
unevictable:0 dirty:27 writeback:0 unstable:0
slab_reclaimable:3967 slab_unreclaimable:4125
mapped:133 shmem:43 pagetables:1674 bounce:0
free:4637 free_pcp:225 free_cma:0
Node 0 active_anon:1698864kB inactive_anon:261256kB active_file:208kB inactive_file:184kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:532kB dirty:108kB writeback:0kB shmem:172kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
DMA free:7316kB min:32kB low:44kB high:56kB active_anon:8064kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:15992kB managed:15908kB mlocked:0kB slab_reclaimable:464kB slab_unreclaimable:40kB kernel_stack:0kB pagetables:24kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
lowmem_reserve[]: 0 992 992 1952
DMA32 free:9088kB min:2048kB low:3064kB high:4080kB active_anon:952176kB inactive_anon:0kB active_file:36kB inactive_file:0kB unevictable:0kB writepending:88kB present:1032192kB managed:1019388kB mlocked:0kB slab_reclaimable:13532kB slab_unreclaimable:16460kB kernel_stack:3552kB pagetables:6672kB bounce:0kB free_pcp:56kB local_pcp:24kB free_cma:0kB
lowmem_reserve[]: 0 0 0 959
Movable free:3644kB min:1980kB low:2960kB high:3940kB active_anon:738560kB inactive_anon:261340kB active_file:188kB inactive_file:640kB unevictable:0kB writepending:20kB present:1048444kB managed:1010816kB mlocked:0kB slab_reclaimable:0kB slab_unreclaimable:0kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:832kB local_pcp:60kB free_cma:0kB
lowmem_reserve[]: 0 0 0 0
DMA: 1*4kB (E) 0*8kB 18*16kB (E) 10*32kB (E) 10*64kB (E) 9*128kB (ME) 8*256kB (E) 2*512kB (E) 2*1024kB (E) 0*2048kB 0*4096kB = 7524kB
DMA32: 417*4kB (UMEH) 181*8kB (UMEH) 68*16kB (UMEH) 48*32kB (UMEH) 14*64kB (MH) 3*128kB (M) 1*256kB (H) 1*512kB (M) 2*1024kB (M) 0*2048kB 0*4096kB = 9836kB
Movable: 1*4kB (M) 1*8kB (M) 1*16kB (M) 1*32kB (M) 0*64kB 1*128kB (M) 2*256kB (M) 4*512kB (M) 1*1024kB (M) 0*2048kB 0*4096kB = 3772kB
378 total pagecache pages
17 pages in swap cache
Swap cache stats: add 17325, delete 17302, find 0/27
Free swap = 978940kB
Total swap = 1048572kB
524157 pages RAM
0 pages HighMem/MovableOnly
12629 pages reserved
0 pages cma reserved
0 pages hwpoisoned
[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name
[ 433] 0 433 4904 5 14 3 82 0 upstart-udev-br
[ 438] 0 438 12371 5 27 3 191 -1000 systemd-udevd
With investigation, skipping page of isolate_lru_pages makes reclaim
void because it returns zero nr_taken easily so LRU shrinking is
effectively nothing and just increases priority aggressively. Finally,
OOM happens.
The problem is that get_scan_count determines nr_to_scan with eligible
zones so although priority drops to zero, it couldn't reclaim any pages
if the LRU contains mostly ineligible pages.
get_scan_count:
size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
size = size >> sc->priority;
Assumes sc->priority is 0 and LRU list is as follows.
N-N-N-N-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H
(Ie, small eligible pages are in the head of LRU but others are
almost ineligible pages)
In that case, size becomes 4 so VM want to scan 4 pages but 4 pages from
tail of the LRU are not eligible pages. If get_scan_count counts
skipped pages, it doesn't reclaim any pages remained after scanning 4
pages so it ends up OOM happening.
This patch makes isolate_lru_pages try to scan pages until it encounters
eligible zones's pages.
[akpm@linux-foundation.org: clean up mind-bending `for' statement. Tweak comment text]
Fixes: 3db65812d688 ("Revert "mm, vmscan: account for skipped pages as a partial scan"")
Link: http://lkml.kernel.org/r/1494457232-27401-1-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-12 22:47:06 +00:00
|
|
|
scan = 0;
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
while (scan < nr_to_scan && !list_empty(src)) {
|
2022-03-22 21:45:41 +00:00
|
|
|
struct list_head *move_to = src;
|
2022-06-17 15:42:45 +00:00
|
|
|
struct folio *folio;
|
2007-07-17 11:03:16 +00:00
|
|
|
|
2022-06-17 15:42:45 +00:00
|
|
|
folio = lru_to_folio(src);
|
|
|
|
prefetchw_prev_lru_folio(folio, src, flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2022-06-17 15:42:45 +00:00
|
|
|
nr_pages = folio_nr_pages(folio);
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
total_scan += nr_pages;
|
|
|
|
|
2022-06-17 15:42:45 +00:00
|
|
|
if (folio_zonenum(folio) > sc->reclaim_idx) {
|
|
|
|
nr_skipped[folio_zonenum(folio)] += nr_pages;
|
|
|
|
move_to = &folios_skipped;
|
2022-03-22 21:45:41 +00:00
|
|
|
goto move;
|
2016-07-28 22:45:37 +00:00
|
|
|
}
|
|
|
|
|
mm: vmscan: scan until it finds eligible pages
Although there are a ton of free swap and anonymous LRU page in elgible
zones, OOM happened.
balloon invoked oom-killer: gfp_mask=0x17080c0(GFP_KERNEL_ACCOUNT|__GFP_ZERO|__GFP_NOTRACK), nodemask=(null), order=0, oom_score_adj=0
CPU: 7 PID: 1138 Comm: balloon Not tainted 4.11.0-rc6-mm1-zram-00289-ge228d67e9677-dirty #17
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
oom_kill_process+0x21d/0x3f0
out_of_memory+0xd8/0x390
__alloc_pages_slowpath+0xbc1/0xc50
__alloc_pages_nodemask+0x1a5/0x1c0
pte_alloc_one+0x20/0x50
__pte_alloc+0x1e/0x110
__handle_mm_fault+0x919/0x960
handle_mm_fault+0x77/0x120
__do_page_fault+0x27a/0x550
trace_do_page_fault+0x43/0x150
do_async_page_fault+0x2c/0x90
async_page_fault+0x28/0x30
Mem-Info:
active_anon:424716 inactive_anon:65314 isolated_anon:0
active_file:52 inactive_file:46 isolated_file:0
unevictable:0 dirty:27 writeback:0 unstable:0
slab_reclaimable:3967 slab_unreclaimable:4125
mapped:133 shmem:43 pagetables:1674 bounce:0
free:4637 free_pcp:225 free_cma:0
Node 0 active_anon:1698864kB inactive_anon:261256kB active_file:208kB inactive_file:184kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:532kB dirty:108kB writeback:0kB shmem:172kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
DMA free:7316kB min:32kB low:44kB high:56kB active_anon:8064kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:15992kB managed:15908kB mlocked:0kB slab_reclaimable:464kB slab_unreclaimable:40kB kernel_stack:0kB pagetables:24kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
lowmem_reserve[]: 0 992 992 1952
DMA32 free:9088kB min:2048kB low:3064kB high:4080kB active_anon:952176kB inactive_anon:0kB active_file:36kB inactive_file:0kB unevictable:0kB writepending:88kB present:1032192kB managed:1019388kB mlocked:0kB slab_reclaimable:13532kB slab_unreclaimable:16460kB kernel_stack:3552kB pagetables:6672kB bounce:0kB free_pcp:56kB local_pcp:24kB free_cma:0kB
lowmem_reserve[]: 0 0 0 959
Movable free:3644kB min:1980kB low:2960kB high:3940kB active_anon:738560kB inactive_anon:261340kB active_file:188kB inactive_file:640kB unevictable:0kB writepending:20kB present:1048444kB managed:1010816kB mlocked:0kB slab_reclaimable:0kB slab_unreclaimable:0kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:832kB local_pcp:60kB free_cma:0kB
lowmem_reserve[]: 0 0 0 0
DMA: 1*4kB (E) 0*8kB 18*16kB (E) 10*32kB (E) 10*64kB (E) 9*128kB (ME) 8*256kB (E) 2*512kB (E) 2*1024kB (E) 0*2048kB 0*4096kB = 7524kB
DMA32: 417*4kB (UMEH) 181*8kB (UMEH) 68*16kB (UMEH) 48*32kB (UMEH) 14*64kB (MH) 3*128kB (M) 1*256kB (H) 1*512kB (M) 2*1024kB (M) 0*2048kB 0*4096kB = 9836kB
Movable: 1*4kB (M) 1*8kB (M) 1*16kB (M) 1*32kB (M) 0*64kB 1*128kB (M) 2*256kB (M) 4*512kB (M) 1*1024kB (M) 0*2048kB 0*4096kB = 3772kB
378 total pagecache pages
17 pages in swap cache
Swap cache stats: add 17325, delete 17302, find 0/27
Free swap = 978940kB
Total swap = 1048572kB
524157 pages RAM
0 pages HighMem/MovableOnly
12629 pages reserved
0 pages cma reserved
0 pages hwpoisoned
[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name
[ 433] 0 433 4904 5 14 3 82 0 upstart-udev-br
[ 438] 0 438 12371 5 27 3 191 -1000 systemd-udevd
With investigation, skipping page of isolate_lru_pages makes reclaim
void because it returns zero nr_taken easily so LRU shrinking is
effectively nothing and just increases priority aggressively. Finally,
OOM happens.
The problem is that get_scan_count determines nr_to_scan with eligible
zones so although priority drops to zero, it couldn't reclaim any pages
if the LRU contains mostly ineligible pages.
get_scan_count:
size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
size = size >> sc->priority;
Assumes sc->priority is 0 and LRU list is as follows.
N-N-N-N-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H
(Ie, small eligible pages are in the head of LRU but others are
almost ineligible pages)
In that case, size becomes 4 so VM want to scan 4 pages but 4 pages from
tail of the LRU are not eligible pages. If get_scan_count counts
skipped pages, it doesn't reclaim any pages remained after scanning 4
pages so it ends up OOM happening.
This patch makes isolate_lru_pages try to scan pages until it encounters
eligible zones's pages.
[akpm@linux-foundation.org: clean up mind-bending `for' statement. Tweak comment text]
Fixes: 3db65812d688 ("Revert "mm, vmscan: account for skipped pages as a partial scan"")
Link: http://lkml.kernel.org/r/1494457232-27401-1-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-12 22:47:06 +00:00
|
|
|
/*
|
2022-06-17 15:42:45 +00:00
|
|
|
* Do not count skipped folios because that makes the function
|
|
|
|
* return with no isolated folios if the LRU mostly contains
|
|
|
|
* ineligible folios. This causes the VM to not reclaim any
|
|
|
|
* folios, triggering a premature OOM.
|
|
|
|
* Account all pages in a folio.
|
mm: vmscan: scan until it finds eligible pages
Although there are a ton of free swap and anonymous LRU page in elgible
zones, OOM happened.
balloon invoked oom-killer: gfp_mask=0x17080c0(GFP_KERNEL_ACCOUNT|__GFP_ZERO|__GFP_NOTRACK), nodemask=(null), order=0, oom_score_adj=0
CPU: 7 PID: 1138 Comm: balloon Not tainted 4.11.0-rc6-mm1-zram-00289-ge228d67e9677-dirty #17
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
oom_kill_process+0x21d/0x3f0
out_of_memory+0xd8/0x390
__alloc_pages_slowpath+0xbc1/0xc50
__alloc_pages_nodemask+0x1a5/0x1c0
pte_alloc_one+0x20/0x50
__pte_alloc+0x1e/0x110
__handle_mm_fault+0x919/0x960
handle_mm_fault+0x77/0x120
__do_page_fault+0x27a/0x550
trace_do_page_fault+0x43/0x150
do_async_page_fault+0x2c/0x90
async_page_fault+0x28/0x30
Mem-Info:
active_anon:424716 inactive_anon:65314 isolated_anon:0
active_file:52 inactive_file:46 isolated_file:0
unevictable:0 dirty:27 writeback:0 unstable:0
slab_reclaimable:3967 slab_unreclaimable:4125
mapped:133 shmem:43 pagetables:1674 bounce:0
free:4637 free_pcp:225 free_cma:0
Node 0 active_anon:1698864kB inactive_anon:261256kB active_file:208kB inactive_file:184kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:532kB dirty:108kB writeback:0kB shmem:172kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
DMA free:7316kB min:32kB low:44kB high:56kB active_anon:8064kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:15992kB managed:15908kB mlocked:0kB slab_reclaimable:464kB slab_unreclaimable:40kB kernel_stack:0kB pagetables:24kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
lowmem_reserve[]: 0 992 992 1952
DMA32 free:9088kB min:2048kB low:3064kB high:4080kB active_anon:952176kB inactive_anon:0kB active_file:36kB inactive_file:0kB unevictable:0kB writepending:88kB present:1032192kB managed:1019388kB mlocked:0kB slab_reclaimable:13532kB slab_unreclaimable:16460kB kernel_stack:3552kB pagetables:6672kB bounce:0kB free_pcp:56kB local_pcp:24kB free_cma:0kB
lowmem_reserve[]: 0 0 0 959
Movable free:3644kB min:1980kB low:2960kB high:3940kB active_anon:738560kB inactive_anon:261340kB active_file:188kB inactive_file:640kB unevictable:0kB writepending:20kB present:1048444kB managed:1010816kB mlocked:0kB slab_reclaimable:0kB slab_unreclaimable:0kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:832kB local_pcp:60kB free_cma:0kB
lowmem_reserve[]: 0 0 0 0
DMA: 1*4kB (E) 0*8kB 18*16kB (E) 10*32kB (E) 10*64kB (E) 9*128kB (ME) 8*256kB (E) 2*512kB (E) 2*1024kB (E) 0*2048kB 0*4096kB = 7524kB
DMA32: 417*4kB (UMEH) 181*8kB (UMEH) 68*16kB (UMEH) 48*32kB (UMEH) 14*64kB (MH) 3*128kB (M) 1*256kB (H) 1*512kB (M) 2*1024kB (M) 0*2048kB 0*4096kB = 9836kB
Movable: 1*4kB (M) 1*8kB (M) 1*16kB (M) 1*32kB (M) 0*64kB 1*128kB (M) 2*256kB (M) 4*512kB (M) 1*1024kB (M) 0*2048kB 0*4096kB = 3772kB
378 total pagecache pages
17 pages in swap cache
Swap cache stats: add 17325, delete 17302, find 0/27
Free swap = 978940kB
Total swap = 1048572kB
524157 pages RAM
0 pages HighMem/MovableOnly
12629 pages reserved
0 pages cma reserved
0 pages hwpoisoned
[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name
[ 433] 0 433 4904 5 14 3 82 0 upstart-udev-br
[ 438] 0 438 12371 5 27 3 191 -1000 systemd-udevd
With investigation, skipping page of isolate_lru_pages makes reclaim
void because it returns zero nr_taken easily so LRU shrinking is
effectively nothing and just increases priority aggressively. Finally,
OOM happens.
The problem is that get_scan_count determines nr_to_scan with eligible
zones so although priority drops to zero, it couldn't reclaim any pages
if the LRU contains mostly ineligible pages.
get_scan_count:
size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
size = size >> sc->priority;
Assumes sc->priority is 0 and LRU list is as follows.
N-N-N-N-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H
(Ie, small eligible pages are in the head of LRU but others are
almost ineligible pages)
In that case, size becomes 4 so VM want to scan 4 pages but 4 pages from
tail of the LRU are not eligible pages. If get_scan_count counts
skipped pages, it doesn't reclaim any pages remained after scanning 4
pages so it ends up OOM happening.
This patch makes isolate_lru_pages try to scan pages until it encounters
eligible zones's pages.
[akpm@linux-foundation.org: clean up mind-bending `for' statement. Tweak comment text]
Fixes: 3db65812d688 ("Revert "mm, vmscan: account for skipped pages as a partial scan"")
Link: http://lkml.kernel.org/r/1494457232-27401-1-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-12 22:47:06 +00:00
|
|
|
*/
|
mm: vmscan: correct some vmscan counters for THP swapout
Commit bd4c82c22c36 ("mm, THP, swap: delay splitting THP after swapped
out"), THP can be swapped out in a whole. But, nr_reclaimed and some
other vm counters still get inc'ed by one even though a whole THP (512
pages) gets swapped out.
This doesn't make too much sense to memory reclaim.
For example, direct reclaim may just need reclaim SWAP_CLUSTER_MAX
pages, reclaiming one THP could fulfill it. But, if nr_reclaimed is not
increased correctly, direct reclaim may just waste time to reclaim more
pages, SWAP_CLUSTER_MAX * 512 pages in worst case.
And, it may cause pgsteal_{kswapd|direct} is greater than
pgscan_{kswapd|direct}, like the below:
pgsteal_kswapd 122933
pgsteal_direct 26600225
pgscan_kswapd 174153
pgscan_direct 14678312
nr_reclaimed and nr_scanned must be fixed in parallel otherwise it would
break some page reclaim logic, e.g.
vmpressure: this looks at the scanned/reclaimed ratio so it won't change
semantics as long as scanned & reclaimed are fixed in parallel.
compaction/reclaim: compaction wants a certain number of physical pages
freed up before going back to compacting.
kswapd priority raising: kswapd raises priority if we scan fewer pages
than the reclaim target (which itself is obviously expressed in order-0
pages). As a result, kswapd can falsely raise its aggressiveness even
when it's making great progress.
Other than nr_scanned and nr_reclaimed, some other counters, e.g.
pgactivate, nr_skipped, nr_ref_keep and nr_unmap_fail need to be fixed too
since they are user visible via cgroup, /proc/vmstat or trace points,
otherwise they would be underreported.
When isolating pages from LRUs, nr_taken has been accounted in base page,
but nr_scanned and nr_skipped are still accounted in THP. It doesn't make
too much sense too since this may cause trace point underreport the
numbers as well.
So accounting those counters in base page instead of accounting THP as one
page.
nr_dirty, nr_unqueued_dirty, nr_congested and nr_writeback are used by
file cache, so they are not impacted by THP swap.
This change may result in lower steal/scan ratio in some cases since THP
may get split during page reclaim, then a part of tail pages get reclaimed
instead of the whole 512 pages, but nr_scanned is accounted by 512,
particularly for direct reclaim. But, this should be not a significant
issue.
Link: http://lkml.kernel.org/r/1559025859-72759-2-git-send-email-yang.shi@linux.alibaba.com
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:59:30 +00:00
|
|
|
scan += nr_pages;
|
2022-03-22 21:45:41 +00:00
|
|
|
|
2022-06-17 15:42:45 +00:00
|
|
|
if (!folio_test_lru(folio))
|
2022-03-22 21:45:41 +00:00
|
|
|
goto move;
|
2022-06-17 15:42:45 +00:00
|
|
|
if (!sc->may_unmap && folio_mapped(folio))
|
2022-03-22 21:45:41 +00:00
|
|
|
goto move;
|
|
|
|
|
2021-02-24 20:08:01 +00:00
|
|
|
/*
|
2022-06-17 15:42:45 +00:00
|
|
|
* Be careful not to clear the lru flag until after we're
|
|
|
|
* sure the folio is not being freed elsewhere -- the
|
|
|
|
* folio release code relies on it.
|
2021-02-24 20:08:01 +00:00
|
|
|
*/
|
2022-06-17 15:42:45 +00:00
|
|
|
if (unlikely(!folio_try_get(folio)))
|
2022-03-22 21:45:41 +00:00
|
|
|
goto move;
|
2007-07-17 11:03:16 +00:00
|
|
|
|
2022-06-17 15:42:45 +00:00
|
|
|
if (!folio_test_clear_lru(folio)) {
|
|
|
|
/* Another thread is already isolating this folio */
|
|
|
|
folio_put(folio);
|
2022-03-22 21:45:41 +00:00
|
|
|
goto move;
|
2007-07-17 11:03:16 +00:00
|
|
|
}
|
2021-02-24 20:08:01 +00:00
|
|
|
|
|
|
|
nr_taken += nr_pages;
|
2022-06-17 15:42:45 +00:00
|
|
|
nr_zone_taken[folio_zonenum(folio)] += nr_pages;
|
2022-03-22 21:45:41 +00:00
|
|
|
move_to = dst;
|
|
|
|
move:
|
2022-06-17 15:42:45 +00:00
|
|
|
list_move(&folio->lru, move_to);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2016-07-28 22:45:37 +00:00
|
|
|
/*
|
2022-06-17 15:42:45 +00:00
|
|
|
* Splice any skipped folios to the start of the LRU list. Note that
|
2016-07-28 22:45:37 +00:00
|
|
|
* this disrupts the LRU order when reclaiming for lower zones but
|
|
|
|
* we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
|
2022-06-17 15:42:45 +00:00
|
|
|
* scanning would soon rescan the same folios to skip and waste lots
|
2022-04-29 06:16:04 +00:00
|
|
|
* of cpu cycles.
|
2016-07-28 22:45:37 +00:00
|
|
|
*/
|
2022-06-17 15:42:45 +00:00
|
|
|
if (!list_empty(&folios_skipped)) {
|
2016-07-28 22:46:59 +00:00
|
|
|
int zid;
|
|
|
|
|
2022-06-17 15:42:45 +00:00
|
|
|
list_splice(&folios_skipped, src);
|
2016-07-28 22:46:59 +00:00
|
|
|
for (zid = 0; zid < MAX_NR_ZONES; zid++) {
|
|
|
|
if (!nr_skipped[zid])
|
|
|
|
continue;
|
|
|
|
|
|
|
|
__count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
|
2017-02-22 23:44:21 +00:00
|
|
|
skipped += nr_skipped[zid];
|
2016-07-28 22:46:59 +00:00
|
|
|
}
|
|
|
|
}
|
mm: vmscan: scan until it finds eligible pages
Although there are a ton of free swap and anonymous LRU page in elgible
zones, OOM happened.
balloon invoked oom-killer: gfp_mask=0x17080c0(GFP_KERNEL_ACCOUNT|__GFP_ZERO|__GFP_NOTRACK), nodemask=(null), order=0, oom_score_adj=0
CPU: 7 PID: 1138 Comm: balloon Not tainted 4.11.0-rc6-mm1-zram-00289-ge228d67e9677-dirty #17
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
Call Trace:
oom_kill_process+0x21d/0x3f0
out_of_memory+0xd8/0x390
__alloc_pages_slowpath+0xbc1/0xc50
__alloc_pages_nodemask+0x1a5/0x1c0
pte_alloc_one+0x20/0x50
__pte_alloc+0x1e/0x110
__handle_mm_fault+0x919/0x960
handle_mm_fault+0x77/0x120
__do_page_fault+0x27a/0x550
trace_do_page_fault+0x43/0x150
do_async_page_fault+0x2c/0x90
async_page_fault+0x28/0x30
Mem-Info:
active_anon:424716 inactive_anon:65314 isolated_anon:0
active_file:52 inactive_file:46 isolated_file:0
unevictable:0 dirty:27 writeback:0 unstable:0
slab_reclaimable:3967 slab_unreclaimable:4125
mapped:133 shmem:43 pagetables:1674 bounce:0
free:4637 free_pcp:225 free_cma:0
Node 0 active_anon:1698864kB inactive_anon:261256kB active_file:208kB inactive_file:184kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:532kB dirty:108kB writeback:0kB shmem:172kB writeback_tmp:0kB unstable:0kB all_unreclaimable? no
DMA free:7316kB min:32kB low:44kB high:56kB active_anon:8064kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:15992kB managed:15908kB mlocked:0kB slab_reclaimable:464kB slab_unreclaimable:40kB kernel_stack:0kB pagetables:24kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
lowmem_reserve[]: 0 992 992 1952
DMA32 free:9088kB min:2048kB low:3064kB high:4080kB active_anon:952176kB inactive_anon:0kB active_file:36kB inactive_file:0kB unevictable:0kB writepending:88kB present:1032192kB managed:1019388kB mlocked:0kB slab_reclaimable:13532kB slab_unreclaimable:16460kB kernel_stack:3552kB pagetables:6672kB bounce:0kB free_pcp:56kB local_pcp:24kB free_cma:0kB
lowmem_reserve[]: 0 0 0 959
Movable free:3644kB min:1980kB low:2960kB high:3940kB active_anon:738560kB inactive_anon:261340kB active_file:188kB inactive_file:640kB unevictable:0kB writepending:20kB present:1048444kB managed:1010816kB mlocked:0kB slab_reclaimable:0kB slab_unreclaimable:0kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:832kB local_pcp:60kB free_cma:0kB
lowmem_reserve[]: 0 0 0 0
DMA: 1*4kB (E) 0*8kB 18*16kB (E) 10*32kB (E) 10*64kB (E) 9*128kB (ME) 8*256kB (E) 2*512kB (E) 2*1024kB (E) 0*2048kB 0*4096kB = 7524kB
DMA32: 417*4kB (UMEH) 181*8kB (UMEH) 68*16kB (UMEH) 48*32kB (UMEH) 14*64kB (MH) 3*128kB (M) 1*256kB (H) 1*512kB (M) 2*1024kB (M) 0*2048kB 0*4096kB = 9836kB
Movable: 1*4kB (M) 1*8kB (M) 1*16kB (M) 1*32kB (M) 0*64kB 1*128kB (M) 2*256kB (M) 4*512kB (M) 1*1024kB (M) 0*2048kB 0*4096kB = 3772kB
378 total pagecache pages
17 pages in swap cache
Swap cache stats: add 17325, delete 17302, find 0/27
Free swap = 978940kB
Total swap = 1048572kB
524157 pages RAM
0 pages HighMem/MovableOnly
12629 pages reserved
0 pages cma reserved
0 pages hwpoisoned
[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name
[ 433] 0 433 4904 5 14 3 82 0 upstart-udev-br
[ 438] 0 438 12371 5 27 3 191 -1000 systemd-udevd
With investigation, skipping page of isolate_lru_pages makes reclaim
void because it returns zero nr_taken easily so LRU shrinking is
effectively nothing and just increases priority aggressively. Finally,
OOM happens.
The problem is that get_scan_count determines nr_to_scan with eligible
zones so although priority drops to zero, it couldn't reclaim any pages
if the LRU contains mostly ineligible pages.
get_scan_count:
size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
size = size >> sc->priority;
Assumes sc->priority is 0 and LRU list is as follows.
N-N-N-N-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H-H
(Ie, small eligible pages are in the head of LRU but others are
almost ineligible pages)
In that case, size becomes 4 so VM want to scan 4 pages but 4 pages from
tail of the LRU are not eligible pages. If get_scan_count counts
skipped pages, it doesn't reclaim any pages remained after scanning 4
pages so it ends up OOM happening.
This patch makes isolate_lru_pages try to scan pages until it encounters
eligible zones's pages.
[akpm@linux-foundation.org: clean up mind-bending `for' statement. Tweak comment text]
Fixes: 3db65812d688 ("Revert "mm, vmscan: account for skipped pages as a partial scan"")
Link: http://lkml.kernel.org/r/1494457232-27401-1-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-12 22:47:06 +00:00
|
|
|
*nr_scanned = total_scan;
|
2017-02-22 23:44:21 +00:00
|
|
|
trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
|
2022-03-22 21:45:41 +00:00
|
|
|
total_scan, skipped, nr_taken,
|
|
|
|
sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru);
|
2017-01-11 00:58:04 +00:00
|
|
|
update_lru_sizes(lruvec, lru, nr_zone_taken);
|
2005-04-16 22:20:36 +00:00
|
|
|
return nr_taken;
|
|
|
|
}
|
|
|
|
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
/**
|
2021-12-24 18:26:22 +00:00
|
|
|
* folio_isolate_lru() - Try to isolate a folio from its LRU list.
|
|
|
|
* @folio: Folio to isolate from its LRU list.
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
*
|
2021-12-24 18:26:22 +00:00
|
|
|
* Isolate a @folio from an LRU list and adjust the vmstat statistic
|
|
|
|
* corresponding to whatever LRU list the folio was on.
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
*
|
2021-12-24 18:26:22 +00:00
|
|
|
* The folio will have its LRU flag cleared. If it was found on the
|
|
|
|
* active list, it will have the Active flag set. If it was found on the
|
|
|
|
* unevictable list, it will have the Unevictable flag set. These flags
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
* may need to be cleared by the caller before letting the page go.
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
*
|
2021-12-24 18:26:22 +00:00
|
|
|
* Context:
|
2018-02-06 23:42:19 +00:00
|
|
|
*
|
2022-09-02 19:45:57 +00:00
|
|
|
* (1) Must be called with an elevated refcount on the folio. This is a
|
|
|
|
* fundamental difference from isolate_lru_folios() (which is called
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
* without a stable reference).
|
2021-12-24 18:26:22 +00:00
|
|
|
* (2) The lru_lock must not be held.
|
|
|
|
* (3) Interrupts must be enabled.
|
|
|
|
*
|
mm: change to return bool for folio_isolate_lru()
Patch series "Change the return value for page isolation functions", v3.
Now the page isolation functions did not return a boolean to indicate
success or not, instead it will return a negative error when failed
to isolate a page. So below code used in most places seem a boolean
success/failure thing, which can confuse people whether the isolation
is successful.
if (folio_isolate_lru(folio))
continue;
Moreover the page isolation functions only return 0 or -EBUSY, and
most users did not care about the negative error except for few users,
thus we can convert all page isolation functions to return a boolean
value, which can remove the confusion to make code more clear.
No functional changes intended in this patch series.
This patch (of 4):
Now the folio_isolate_lru() did not return a boolean value to indicate
isolation success or not, however below code checking the return value can
make people think that it was a boolean success/failure thing, which makes
people easy to make mistakes (see the fix patch[1]).
if (folio_isolate_lru(folio))
continue;
Thus it's better to check the negative error value expilictly returned by
folio_isolate_lru(), which makes code more clear per Linus's
suggestion[2]. Moreover Matthew suggested we can convert the isolation
functions to return a boolean[3], since most users did not care about the
negative error value, and can also remove the confusing of checking return
value.
So this patch converts the folio_isolate_lru() to return a boolean value,
which means return 'true' to indicate the folio isolation is successful,
and 'false' means a failure to isolation. Meanwhile changing all users'
logic of checking the isolation state.
No functional changes intended.
[1] https://lore.kernel.org/all/20230131063206.28820-1-Kuan-Ying.Lee@mediatek.com/T/#u
[2] https://lore.kernel.org/all/CAHk-=wiBrY+O-4=2mrbVyxR+hOqfdJ=Do6xoucfJ9_5az01L4Q@mail.gmail.com/
[3] https://lore.kernel.org/all/Y+sTFqwMNAjDvxw3@casper.infradead.org/
Link: https://lkml.kernel.org/r/cover.1676424378.git.baolin.wang@linux.alibaba.com
Link: https://lkml.kernel.org/r/8a4e3679ed4196168efadf7ea36c038f2f7d5aa9.1676424378.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: SeongJae Park <sj@kernel.org>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-15 10:39:34 +00:00
|
|
|
* Return: true if the folio was removed from an LRU list.
|
|
|
|
* false if the folio was not on an LRU list.
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
*/
|
mm: change to return bool for folio_isolate_lru()
Patch series "Change the return value for page isolation functions", v3.
Now the page isolation functions did not return a boolean to indicate
success or not, instead it will return a negative error when failed
to isolate a page. So below code used in most places seem a boolean
success/failure thing, which can confuse people whether the isolation
is successful.
if (folio_isolate_lru(folio))
continue;
Moreover the page isolation functions only return 0 or -EBUSY, and
most users did not care about the negative error except for few users,
thus we can convert all page isolation functions to return a boolean
value, which can remove the confusion to make code more clear.
No functional changes intended in this patch series.
This patch (of 4):
Now the folio_isolate_lru() did not return a boolean value to indicate
isolation success or not, however below code checking the return value can
make people think that it was a boolean success/failure thing, which makes
people easy to make mistakes (see the fix patch[1]).
if (folio_isolate_lru(folio))
continue;
Thus it's better to check the negative error value expilictly returned by
folio_isolate_lru(), which makes code more clear per Linus's
suggestion[2]. Moreover Matthew suggested we can convert the isolation
functions to return a boolean[3], since most users did not care about the
negative error value, and can also remove the confusing of checking return
value.
So this patch converts the folio_isolate_lru() to return a boolean value,
which means return 'true' to indicate the folio isolation is successful,
and 'false' means a failure to isolation. Meanwhile changing all users'
logic of checking the isolation state.
No functional changes intended.
[1] https://lore.kernel.org/all/20230131063206.28820-1-Kuan-Ying.Lee@mediatek.com/T/#u
[2] https://lore.kernel.org/all/CAHk-=wiBrY+O-4=2mrbVyxR+hOqfdJ=Do6xoucfJ9_5az01L4Q@mail.gmail.com/
[3] https://lore.kernel.org/all/Y+sTFqwMNAjDvxw3@casper.infradead.org/
Link: https://lkml.kernel.org/r/cover.1676424378.git.baolin.wang@linux.alibaba.com
Link: https://lkml.kernel.org/r/8a4e3679ed4196168efadf7ea36c038f2f7d5aa9.1676424378.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: SeongJae Park <sj@kernel.org>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-15 10:39:34 +00:00
|
|
|
bool folio_isolate_lru(struct folio *folio)
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
{
|
mm: change to return bool for folio_isolate_lru()
Patch series "Change the return value for page isolation functions", v3.
Now the page isolation functions did not return a boolean to indicate
success or not, instead it will return a negative error when failed
to isolate a page. So below code used in most places seem a boolean
success/failure thing, which can confuse people whether the isolation
is successful.
if (folio_isolate_lru(folio))
continue;
Moreover the page isolation functions only return 0 or -EBUSY, and
most users did not care about the negative error except for few users,
thus we can convert all page isolation functions to return a boolean
value, which can remove the confusion to make code more clear.
No functional changes intended in this patch series.
This patch (of 4):
Now the folio_isolate_lru() did not return a boolean value to indicate
isolation success or not, however below code checking the return value can
make people think that it was a boolean success/failure thing, which makes
people easy to make mistakes (see the fix patch[1]).
if (folio_isolate_lru(folio))
continue;
Thus it's better to check the negative error value expilictly returned by
folio_isolate_lru(), which makes code more clear per Linus's
suggestion[2]. Moreover Matthew suggested we can convert the isolation
functions to return a boolean[3], since most users did not care about the
negative error value, and can also remove the confusing of checking return
value.
So this patch converts the folio_isolate_lru() to return a boolean value,
which means return 'true' to indicate the folio isolation is successful,
and 'false' means a failure to isolation. Meanwhile changing all users'
logic of checking the isolation state.
No functional changes intended.
[1] https://lore.kernel.org/all/20230131063206.28820-1-Kuan-Ying.Lee@mediatek.com/T/#u
[2] https://lore.kernel.org/all/CAHk-=wiBrY+O-4=2mrbVyxR+hOqfdJ=Do6xoucfJ9_5az01L4Q@mail.gmail.com/
[3] https://lore.kernel.org/all/Y+sTFqwMNAjDvxw3@casper.infradead.org/
Link: https://lkml.kernel.org/r/cover.1676424378.git.baolin.wang@linux.alibaba.com
Link: https://lkml.kernel.org/r/8a4e3679ed4196168efadf7ea36c038f2f7d5aa9.1676424378.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: SeongJae Park <sj@kernel.org>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-15 10:39:34 +00:00
|
|
|
bool ret = false;
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
|
2021-12-24 18:26:22 +00:00
|
|
|
VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
|
2011-05-25 00:12:21 +00:00
|
|
|
|
2021-12-24 18:26:22 +00:00
|
|
|
if (folio_test_clear_lru(folio)) {
|
2012-05-29 22:07:09 +00:00
|
|
|
struct lruvec *lruvec;
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
|
2021-12-24 18:26:22 +00:00
|
|
|
folio_get(folio);
|
2021-06-29 01:59:47 +00:00
|
|
|
lruvec = folio_lruvec_lock_irq(folio);
|
2021-12-24 18:26:22 +00:00
|
|
|
lruvec_del_folio(lruvec, folio);
|
2020-12-15 20:34:29 +00:00
|
|
|
unlock_page_lruvec_irq(lruvec);
|
mm: change to return bool for folio_isolate_lru()
Patch series "Change the return value for page isolation functions", v3.
Now the page isolation functions did not return a boolean to indicate
success or not, instead it will return a negative error when failed
to isolate a page. So below code used in most places seem a boolean
success/failure thing, which can confuse people whether the isolation
is successful.
if (folio_isolate_lru(folio))
continue;
Moreover the page isolation functions only return 0 or -EBUSY, and
most users did not care about the negative error except for few users,
thus we can convert all page isolation functions to return a boolean
value, which can remove the confusion to make code more clear.
No functional changes intended in this patch series.
This patch (of 4):
Now the folio_isolate_lru() did not return a boolean value to indicate
isolation success or not, however below code checking the return value can
make people think that it was a boolean success/failure thing, which makes
people easy to make mistakes (see the fix patch[1]).
if (folio_isolate_lru(folio))
continue;
Thus it's better to check the negative error value expilictly returned by
folio_isolate_lru(), which makes code more clear per Linus's
suggestion[2]. Moreover Matthew suggested we can convert the isolation
functions to return a boolean[3], since most users did not care about the
negative error value, and can also remove the confusing of checking return
value.
So this patch converts the folio_isolate_lru() to return a boolean value,
which means return 'true' to indicate the folio isolation is successful,
and 'false' means a failure to isolation. Meanwhile changing all users'
logic of checking the isolation state.
No functional changes intended.
[1] https://lore.kernel.org/all/20230131063206.28820-1-Kuan-Ying.Lee@mediatek.com/T/#u
[2] https://lore.kernel.org/all/CAHk-=wiBrY+O-4=2mrbVyxR+hOqfdJ=Do6xoucfJ9_5az01L4Q@mail.gmail.com/
[3] https://lore.kernel.org/all/Y+sTFqwMNAjDvxw3@casper.infradead.org/
Link: https://lkml.kernel.org/r/cover.1676424378.git.baolin.wang@linux.alibaba.com
Link: https://lkml.kernel.org/r/8a4e3679ed4196168efadf7ea36c038f2f7d5aa9.1676424378.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: SeongJae Park <sj@kernel.org>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-15 10:39:34 +00:00
|
|
|
ret = true;
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
}
|
mm/lru: introduce TestClearPageLRU()
Currently lru_lock still guards both lru list and page's lru bit, that's
ok. but if we want to use specific lruvec lock on the page, we need to
pin down the page's lruvec/memcg during locking. Just taking lruvec lock
first may be undermined by the page's memcg charge/migration. To fix this
problem, we will clear the lru bit out of locking and use it as pin down
action to block the page isolation in memcg changing.
So now a standard steps of page isolation is following:
1, get_page(); #pin the page avoid to be free
2, TestClearPageLRU(); #block other isolation like memcg change
3, spin_lock on lru_lock; #serialize lru list access
4, delete page from lru list;
This patch start with the first part: TestClearPageLRU, which combines
PageLRU check and ClearPageLRU into a macro func TestClearPageLRU. This
function will be used as page isolation precondition to prevent other
isolations some where else. Then there are may !PageLRU page on lru list,
need to remove BUG() checking accordingly.
There 2 rules for lru bit now:
1, the lru bit still indicate if a page on lru list, just in some
temporary moment(isolating), the page may have no lru bit when
it's on lru list. but the page still must be on lru list when the
lru bit set.
2, have to remove lru bit before delete it from lru list.
As Andrew Morton mentioned this change would dirty cacheline for a page
which isn't on the LRU. But the loss would be acceptable in Rong Chen
<rong.a.chen@intel.com> report:
https://lore.kernel.org/lkml/20200304090301.GB5972@shao2-debian/
Link: https://lkml.kernel.org/r/1604566549-62481-15-git-send-email-alex.shi@linux.alibaba.com
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 20:34:16 +00:00
|
|
|
|
vmscan: move isolate_lru_page() to vmscan.c
On large memory systems, the VM can spend way too much time scanning
through pages that it cannot (or should not) evict from memory. Not only
does it use up CPU time, but it also provokes lock contention and can
leave large systems under memory presure in a catatonic state.
This patch series improves VM scalability by:
1) putting filesystem backed, swap backed and unevictable pages
onto their own LRUs, so the system only scans the pages that it
can/should evict from memory
2) switching to two handed clock replacement for the anonymous LRUs,
so the number of pages that need to be scanned when the system
starts swapping is bound to a reasonable number
3) keeping unevictable pages off the LRU completely, so the
VM does not waste CPU time scanning them. ramfs, ramdisk,
SHM_LOCKED shared memory segments and mlock()ed VMA pages
are keept on the unevictable list.
This patch:
isolate_lru_page logically belongs to be in vmscan.c than migrate.c.
It is tough, because we don't need that function without memory migration
so there is a valid argument to have it in migrate.c. However a
subsequent patch needs to make use of it in the core mm, so we can happily
move it to vmscan.c.
Also, make the function a little more generic by not requiring that it
adds an isolated page to a given list. Callers can do that.
Note that we now have '__isolate_lru_page()', that does
something quite different, visible outside of vmscan.c
for use with memory controller. Methinks we need to
rationalize these names/purposes. --lts
[akpm@linux-foundation.org: fix mm/memory_hotplug.c build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:09 +00:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2009-09-22 00:01:38 +00:00
|
|
|
/*
|
2012-12-18 22:23:28 +00:00
|
|
|
* A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
|
2019-12-01 01:56:05 +00:00
|
|
|
* then get rescheduled. When there are massive number of tasks doing page
|
2012-12-18 22:23:28 +00:00
|
|
|
* allocation, such sleeping direct reclaimers may keep piling up on each CPU,
|
|
|
|
* the LRU list will go small and be scanned faster than necessary, leading to
|
|
|
|
* unnecessary swapping, thrashing and OOM.
|
2009-09-22 00:01:38 +00:00
|
|
|
*/
|
2016-07-28 22:45:31 +00:00
|
|
|
static int too_many_isolated(struct pglist_data *pgdat, int file,
|
2009-09-22 00:01:38 +00:00
|
|
|
struct scan_control *sc)
|
|
|
|
{
|
|
|
|
unsigned long inactive, isolated;
|
2021-11-05 20:42:29 +00:00
|
|
|
bool too_many;
|
2009-09-22 00:01:38 +00:00
|
|
|
|
|
|
|
if (current_is_kswapd())
|
|
|
|
return 0;
|
|
|
|
|
2019-12-01 01:55:40 +00:00
|
|
|
if (!writeback_throttling_sane(sc))
|
2009-09-22 00:01:38 +00:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (file) {
|
2016-07-28 22:45:31 +00:00
|
|
|
inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
|
|
|
|
isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
|
2009-09-22 00:01:38 +00:00
|
|
|
} else {
|
2016-07-28 22:45:31 +00:00
|
|
|
inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
|
|
|
|
isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
|
2009-09-22 00:01:38 +00:00
|
|
|
}
|
|
|
|
|
mm/vmscan.c: avoid possible deadlock caused by too_many_isolated()
Neil found that if too_many_isolated() returns true while performing
direct reclaim we can end up waiting for other threads to complete their
direct reclaim. If those threads are allowed to enter the FS or IO to
free memory, but this thread is not, then it is possible that those
threads will be waiting on this thread and so we get a circular deadlock.
some task enters direct reclaim with GFP_KERNEL
=> too_many_isolated() false
=> vmscan and run into dirty pages
=> pageout()
=> take some FS lock
=> fs/block code does GFP_NOIO allocation
=> enter direct reclaim again
=> too_many_isolated() true
=> waiting for others to progress, however the other
tasks may be circular waiting for the FS lock..
The fix is to let !__GFP_IO and !__GFP_FS direct reclaims enjoy higher
priority than normal ones, by lowering the throttle threshold for the
latter.
Allowing ~1/8 isolated pages in normal is large enough. For example, for
a 1GB LRU list, that's ~128MB isolated pages, or 1k blocked tasks (each
isolates 32 4KB pages), or 64 blocked tasks per logical CPU (assuming 16
logical CPUs per NUMA node). So it's not likely some CPU goes idle
waiting (when it could make progress) because of this limit: there are
much more sleeping reclaim tasks than the number of CPU, so the task may
well be blocked by some low level queue/lock anyway.
Now !GFP_IOFS reclaims won't be waiting for GFP_IOFS reclaims to progress.
They will be blocked only when there are too many concurrent !GFP_IOFS
reclaims, however that's very unlikely because the IO-less direct reclaims
is able to progress much more faster, and they won't deadlock each other.
The threshold is raised high enough for them, so that there can be
sufficient parallel progress of !GFP_IOFS reclaims.
[akpm@linux-foundation.org: tweak comment]
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: Torsten Kaiser <just.for.lkml@googlemail.com>
Tested-by: NeilBrown <neilb@suse.de>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-18 22:23:31 +00:00
|
|
|
/*
|
|
|
|
* GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
|
|
|
|
* won't get blocked by normal direct-reclaimers, forming a circular
|
|
|
|
* deadlock.
|
|
|
|
*/
|
2015-11-07 00:28:21 +00:00
|
|
|
if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
|
mm/vmscan.c: avoid possible deadlock caused by too_many_isolated()
Neil found that if too_many_isolated() returns true while performing
direct reclaim we can end up waiting for other threads to complete their
direct reclaim. If those threads are allowed to enter the FS or IO to
free memory, but this thread is not, then it is possible that those
threads will be waiting on this thread and so we get a circular deadlock.
some task enters direct reclaim with GFP_KERNEL
=> too_many_isolated() false
=> vmscan and run into dirty pages
=> pageout()
=> take some FS lock
=> fs/block code does GFP_NOIO allocation
=> enter direct reclaim again
=> too_many_isolated() true
=> waiting for others to progress, however the other
tasks may be circular waiting for the FS lock..
The fix is to let !__GFP_IO and !__GFP_FS direct reclaims enjoy higher
priority than normal ones, by lowering the throttle threshold for the
latter.
Allowing ~1/8 isolated pages in normal is large enough. For example, for
a 1GB LRU list, that's ~128MB isolated pages, or 1k blocked tasks (each
isolates 32 4KB pages), or 64 blocked tasks per logical CPU (assuming 16
logical CPUs per NUMA node). So it's not likely some CPU goes idle
waiting (when it could make progress) because of this limit: there are
much more sleeping reclaim tasks than the number of CPU, so the task may
well be blocked by some low level queue/lock anyway.
Now !GFP_IOFS reclaims won't be waiting for GFP_IOFS reclaims to progress.
They will be blocked only when there are too many concurrent !GFP_IOFS
reclaims, however that's very unlikely because the IO-less direct reclaims
is able to progress much more faster, and they won't deadlock each other.
The threshold is raised high enough for them, so that there can be
sufficient parallel progress of !GFP_IOFS reclaims.
[akpm@linux-foundation.org: tweak comment]
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: Torsten Kaiser <just.for.lkml@googlemail.com>
Tested-by: NeilBrown <neilb@suse.de>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-18 22:23:31 +00:00
|
|
|
inactive >>= 3;
|
|
|
|
|
2021-11-05 20:42:29 +00:00
|
|
|
too_many = isolated > inactive;
|
|
|
|
|
|
|
|
/* Wake up tasks throttled due to too_many_isolated. */
|
|
|
|
if (!too_many)
|
|
|
|
wake_throttle_isolated(pgdat);
|
|
|
|
|
|
|
|
return too_many;
|
2009-09-22 00:01:38 +00:00
|
|
|
}
|
|
|
|
|
2019-05-14 00:17:00 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* move_folios_to_lru() moves folios from private @list to appropriate LRU list.
|
2022-06-17 15:42:46 +00:00
|
|
|
* On return, @list is reused as a list of folios to be freed by the caller.
|
2019-05-14 00:17:00 +00:00
|
|
|
*
|
|
|
|
* Returns the number of pages moved to the given lruvec.
|
|
|
|
*/
|
2022-09-02 19:45:57 +00:00
|
|
|
static unsigned int move_folios_to_lru(struct lruvec *lruvec,
|
|
|
|
struct list_head *list)
|
2010-08-10 00:19:30 +00:00
|
|
|
{
|
2019-05-14 00:17:00 +00:00
|
|
|
int nr_pages, nr_moved = 0;
|
2022-06-17 15:42:46 +00:00
|
|
|
LIST_HEAD(folios_to_free);
|
2010-08-10 00:19:30 +00:00
|
|
|
|
2019-05-14 00:17:00 +00:00
|
|
|
while (!list_empty(list)) {
|
2022-06-17 15:42:46 +00:00
|
|
|
struct folio *folio = lru_to_folio(list);
|
|
|
|
|
|
|
|
VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
|
|
|
|
list_del(&folio->lru);
|
|
|
|
if (unlikely(!folio_evictable(folio))) {
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
2022-06-17 15:42:46 +00:00
|
|
|
folio_putback_lru(folio);
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
2010-08-10 00:19:30 +00:00
|
|
|
continue;
|
|
|
|
}
|
2012-05-29 22:07:09 +00:00
|
|
|
|
2020-12-15 20:33:37 +00:00
|
|
|
/*
|
2022-06-17 15:42:46 +00:00
|
|
|
* The folio_set_lru needs to be kept here for list integrity.
|
2020-12-15 20:33:37 +00:00
|
|
|
* Otherwise:
|
2022-09-02 19:45:57 +00:00
|
|
|
* #0 move_folios_to_lru #1 release_pages
|
2022-06-17 15:42:46 +00:00
|
|
|
* if (!folio_put_testzero())
|
|
|
|
* if (folio_put_testzero())
|
|
|
|
* !lru //skip lru_lock
|
|
|
|
* folio_set_lru()
|
|
|
|
* list_add(&folio->lru,)
|
|
|
|
* list_add(&folio->lru,)
|
2020-12-15 20:33:37 +00:00
|
|
|
*/
|
2022-06-17 15:42:46 +00:00
|
|
|
folio_set_lru(folio);
|
2019-05-14 00:17:00 +00:00
|
|
|
|
2022-06-17 15:42:46 +00:00
|
|
|
if (unlikely(folio_put_testzero(folio))) {
|
|
|
|
__folio_clear_lru_flags(folio);
|
2012-01-13 01:19:56 +00:00
|
|
|
|
2022-06-17 15:42:46 +00:00
|
|
|
if (unlikely(folio_test_large(folio))) {
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
2022-06-17 17:50:17 +00:00
|
|
|
destroy_large_folio(folio);
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
2012-01-13 01:19:56 +00:00
|
|
|
} else
|
2022-06-17 15:42:46 +00:00
|
|
|
list_add(&folio->lru, &folios_to_free);
|
2020-12-15 20:33:37 +00:00
|
|
|
|
|
|
|
continue;
|
2010-08-10 00:19:30 +00:00
|
|
|
}
|
2020-12-15 20:33:37 +00:00
|
|
|
|
2020-12-15 20:34:02 +00:00
|
|
|
/*
|
|
|
|
* All pages were isolated from the same lruvec (and isolation
|
|
|
|
* inhibits memcg migration).
|
|
|
|
*/
|
2022-06-17 15:42:46 +00:00
|
|
|
VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
|
|
|
|
lruvec_add_folio(lruvec, folio);
|
|
|
|
nr_pages = folio_nr_pages(folio);
|
2020-12-15 20:33:37 +00:00
|
|
|
nr_moved += nr_pages;
|
2022-06-17 15:42:46 +00:00
|
|
|
if (folio_test_active(folio))
|
2020-12-15 20:33:37 +00:00
|
|
|
workingset_age_nonresident(lruvec, nr_pages);
|
2010-08-10 00:19:30 +00:00
|
|
|
}
|
|
|
|
|
2012-01-13 01:20:07 +00:00
|
|
|
/*
|
|
|
|
* To save our caller's stack, now use input list for pages to free.
|
|
|
|
*/
|
2022-06-17 15:42:46 +00:00
|
|
|
list_splice(&folios_to_free, list);
|
2019-05-14 00:17:00 +00:00
|
|
|
|
|
|
|
return nr_moved;
|
2010-08-10 00:19:30 +00:00
|
|
|
}
|
|
|
|
|
2014-06-04 23:07:42 +00:00
|
|
|
/*
|
2022-04-29 06:16:04 +00:00
|
|
|
* If a kernel thread (such as nfsd for loop-back mounts) services a backing
|
|
|
|
* device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
|
|
|
|
* we should not throttle. Otherwise it is safe to do so.
|
2014-06-04 23:07:42 +00:00
|
|
|
*/
|
|
|
|
static int current_may_throttle(void)
|
|
|
|
{
|
2022-03-22 21:39:10 +00:00
|
|
|
return !(current->flags & PF_LOCAL_THROTTLE);
|
2014-06-04 23:07:42 +00:00
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2016-07-28 22:45:37 +00:00
|
|
|
* shrink_inactive_list() is a helper for shrink_node(). It returns the number
|
[PATCH] vmscan: rename functions
We have:
try_to_free_pages
->shrink_caches(struct zone **zones, ..)
->shrink_zone(struct zone *, ...)
->shrink_cache(struct zone *, ...)
->shrink_list(struct list_head *, ...)
->refill_inactive_list((struct zone *, ...)
which is fairly irrational.
Rename things so that we have
try_to_free_pages
->shrink_zones(struct zone **zones, ..)
->shrink_zone(struct zone *, ...)
->shrink_inactive_list(struct zone *, ...)
->shrink_page_list(struct list_head *, ...)
->shrink_active_list(struct zone *, ...)
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Christoph Lameter <christoph@lameter.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-22 08:08:21 +00:00
|
|
|
* of reclaimed pages
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2022-09-02 19:45:57 +00:00
|
|
|
static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
|
|
|
|
struct lruvec *lruvec, struct scan_control *sc,
|
|
|
|
enum lru_list lru)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2022-09-02 19:45:57 +00:00
|
|
|
LIST_HEAD(folio_list);
|
2010-08-10 00:19:28 +00:00
|
|
|
unsigned long nr_scanned;
|
2020-06-03 23:01:18 +00:00
|
|
|
unsigned int nr_reclaimed = 0;
|
2010-08-10 00:19:28 +00:00
|
|
|
unsigned long nr_taken;
|
2019-03-05 23:48:15 +00:00
|
|
|
struct reclaim_stat stat;
|
2020-06-03 23:02:34 +00:00
|
|
|
bool file = is_file_lru(lru);
|
2019-05-14 00:22:33 +00:00
|
|
|
enum vm_event_item item;
|
2016-07-28 22:45:31 +00:00
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(lruvec);
|
2017-09-06 23:21:11 +00:00
|
|
|
bool stalled = false;
|
2009-06-16 22:31:40 +00:00
|
|
|
|
2016-07-28 22:45:31 +00:00
|
|
|
while (unlikely(too_many_isolated(pgdat, file, sc))) {
|
2017-09-06 23:21:11 +00:00
|
|
|
if (stalled)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
/* wait a bit for the reclaimer. */
|
|
|
|
stalled = true;
|
2021-11-05 20:42:42 +00:00
|
|
|
reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
|
2009-09-22 00:01:38 +00:00
|
|
|
|
|
|
|
/* We are about to die and free our memory. Return now. */
|
|
|
|
if (fatal_signal_pending(current))
|
|
|
|
return SWAP_CLUSTER_MAX;
|
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
lru_add_drain();
|
2011-11-01 00:06:55 +00:00
|
|
|
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
2009-09-22 00:01:36 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
|
2019-03-05 23:46:55 +00:00
|
|
|
&nr_scanned, sc, lru);
|
2012-05-29 22:06:59 +00:00
|
|
|
|
2016-07-28 22:45:31 +00:00
|
|
|
__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
|
2022-10-26 18:01:33 +00:00
|
|
|
item = PGSCAN_KSWAPD + reclaimer_offset();
|
2019-12-01 01:55:40 +00:00
|
|
|
if (!cgroup_reclaim(sc))
|
2019-05-14 00:22:33 +00:00
|
|
|
__count_vm_events(item, nr_scanned);
|
|
|
|
__count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
|
2020-06-03 23:02:34 +00:00
|
|
|
__count_vm_events(PGSCAN_ANON + file, nr_scanned);
|
|
|
|
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
2009-09-22 00:01:36 +00:00
|
|
|
|
2012-03-21 23:34:02 +00:00
|
|
|
if (nr_taken == 0)
|
2010-08-10 00:19:30 +00:00
|
|
|
return 0;
|
2007-07-17 11:03:16 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
|
2007-08-22 21:01:26 +00:00
|
|
|
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
2022-09-02 19:45:57 +00:00
|
|
|
move_folios_to_lru(lruvec, &folio_list);
|
2020-06-03 23:02:34 +00:00
|
|
|
|
|
|
|
__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
|
2022-10-26 18:01:33 +00:00
|
|
|
item = PGSTEAL_KSWAPD + reclaimer_offset();
|
2019-12-01 01:55:40 +00:00
|
|
|
if (!cgroup_reclaim(sc))
|
2019-05-14 00:22:33 +00:00
|
|
|
__count_vm_events(item, nr_reclaimed);
|
|
|
|
__count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
|
2020-06-03 23:02:34 +00:00
|
|
|
__count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
2012-01-13 01:20:07 +00:00
|
|
|
|
mm: vmscan: make rotations a secondary factor in balancing anon vs file
We noticed a 2% webserver throughput regression after upgrading from 5.6.
This could be tracked down to a shift in the anon/file reclaim balance
(confirmed with swappiness) that resulted in worse reclaim efficiency and
thus more kswapd activity for the same outcome.
The change that exposed the problem is aae466b0052e ("mm/swap: implement
workingset detection for anonymous LRU"). By qualifying swapins based on
their refault distance, it lowered the cost of anon reclaim in this
workload, in turn causing (much) more anon scanning than before. Scanning
the anon list is more expensive due to the higher ratio of mmapped pages
that may rotate during reclaim, and so the result was an increase in %sys
time.
Right now, rotations aren't considered a cost when balancing scan pressure
between LRUs. We can end up with very few file refaults putting all the
scan pressure on hot anon pages that are rotated en masse, don't get
reclaimed, and never push back on the file LRU again. We still only
reclaim file cache in that case, but we burn a lot CPU rotating anon
pages. It's "fair" from an LRU age POV, but doesn't reflect the real cost
it imposes on the system.
Consider rotations as a secondary factor in balancing the LRUs. This
doesn't attempt to make a precise comparison between IO cost and CPU cost,
it just says: if reloads are about comparable between the lists, or
rotations are overwhelmingly different, adjust for CPU work.
This fixed the regression on our webservers. It has since been deployed
to the entire Meta fleet and hasn't caused any problems.
Link: https://lkml.kernel.org/r/20221013193113.726425-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@surriel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-10-13 19:31:13 +00:00
|
|
|
lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed);
|
2022-09-02 19:45:57 +00:00
|
|
|
mem_cgroup_uncharge_list(&folio_list);
|
|
|
|
free_unref_page_list(&folio_list);
|
tracing, vmscan: add trace events for LRU list shrinking
There have been numerous reports of stalls that pointed at the problem
being somewhere in the VM. There are multiple roots to the problems which
means dealing with any of the root problems in isolation is tricky to
justify on their own and they would still need integration testing. This
patch series puts together two different patch sets which in combination
should tackle some of the root causes of latency problems being reported.
Patch 1 adds a tracepoint for shrink_inactive_list. For this series, the
most important results is being able to calculate the scanning/reclaim
ratio as a measure of the amount of work being done by page reclaim.
Patch 2 accounts for time spent in congestion_wait.
Patches 3-6 were originally developed by Kosaki Motohiro but reworked for
this series. It has been noted that lumpy reclaim is far too aggressive
and trashes the system somewhat. As SLUB uses high-order allocations, a
large cost incurred by lumpy reclaim will be noticeable. It was also
reported during transparent hugepage support testing that lumpy reclaim
was trashing the system and these patches should mitigate that problem
without disabling lumpy reclaim.
Patch 7 adds wait_iff_congested() and replaces some callers of
congestion_wait(). wait_iff_congested() only sleeps if there is a BDI
that is currently congested. Patch 8 notes that any BDI being congested
is not necessarily a problem because there could be multiple BDIs of
varying speeds and numberous zones. It attempts to track when a zone
being reclaimed contains many pages backed by a congested BDI and if so,
reclaimers wait on the congestion queue.
I ran a number of tests with monitoring on X86, X86-64 and PPC64. Each
machine had 3G of RAM and the CPUs were
X86: Intel P4 2-core
X86-64: AMD Phenom 4-core
PPC64: PPC970MP
Each used a single disk and the onboard IO controller. Dirty ratio was
left at 20. I'm just going to report for X86-64 and PPC64 in a vague
attempt to keep this report short. Four kernels were tested each based on
v2.6.36-rc4
traceonly-v2r2: Patches 1 and 2 to instrument vmscan reclaims and congestion_wait
lowlumpy-v2r3: Patches 1-6 to test if lumpy reclaim is better
waitcongest-v2r3: Patches 1-7 to only wait on congestion
waitwriteback-v2r4: Patches 1-8 to detect when a zone is congested
nocongest-v1r5: Patches 1-3 for testing wait_iff_congestion
nodirect-v1r5: Patches 1-10 to disable filesystem writeback for better IO
The tests run were as follows
kernbench
compile-based benchmark. Smoke test performance
sysbench
OLTP read-only benchmark. Will be re-run in the future as read-write
micro-mapped-file-stream
This is a micro-benchmark from Johannes Weiner that accesses a
large sparse-file through mmap(). It was configured to run in only
single-CPU mode but can be indicative of how well page reclaim
identifies suitable pages.
stress-highalloc
Tries to allocate huge pages under heavy load.
kernbench, iozone and sysbench did not report any performance regression
on any machine. sysbench did pressure the system lightly and there was
reclaim activity but there were no difference of major interest between
the kernels.
X86-64 micro-mapped-file-stream
traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4
pgalloc_dma 1639.00 ( 0.00%) 667.00 (-145.73%) 1167.00 ( -40.45%) 578.00 (-183.56%)
pgalloc_dma32 2842410.00 ( 0.00%) 2842626.00 ( 0.01%) 2843043.00 ( 0.02%) 2843014.00 ( 0.02%)
pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
pgsteal_dma 729.00 ( 0.00%) 85.00 (-757.65%) 609.00 ( -19.70%) 125.00 (-483.20%)
pgsteal_dma32 2338721.00 ( 0.00%) 2447354.00 ( 4.44%) 2429536.00 ( 3.74%) 2436772.00 ( 4.02%)
pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
pgscan_kswapd_dma 1469.00 ( 0.00%) 532.00 (-176.13%) 1078.00 ( -36.27%) 220.00 (-567.73%)
pgscan_kswapd_dma32 4597713.00 ( 0.00%) 4503597.00 ( -2.09%) 4295673.00 ( -7.03%) 3891686.00 ( -18.14%)
pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
pgscan_direct_dma 71.00 ( 0.00%) 134.00 ( 47.01%) 243.00 ( 70.78%) 352.00 ( 79.83%)
pgscan_direct_dma32 305820.00 ( 0.00%) 280204.00 ( -9.14%) 600518.00 ( 49.07%) 957485.00 ( 68.06%)
pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
pageoutrun 16296.00 ( 0.00%) 21254.00 ( 23.33%) 18447.00 ( 11.66%) 20067.00 ( 18.79%)
allocstall 443.00 ( 0.00%) 273.00 ( -62.27%) 513.00 ( 13.65%) 1568.00 ( 71.75%)
These are based on the raw figures taken from /proc/vmstat. It's a rough
measure of reclaim activity. Note that allocstall counts are higher
because we are entering direct reclaim more often as a result of not
sleeping in congestion. In itself, it's not necessarily a bad thing.
It's easier to get a view of what happened from the vmscan tracepoint
report.
FTrace Reclaim Statistics: vmscan
traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3 waitwriteback-v2r4
Direct reclaims 443 273 513 1568
Direct reclaim pages scanned 305968 280402 600825 957933
Direct reclaim pages reclaimed 43503 19005 30327 117191
Direct reclaim write file async I/O 0 0 0 0
Direct reclaim write anon async I/O 0 3 4 12
Direct reclaim write file sync I/O 0 0 0 0
Direct reclaim write anon sync I/O 0 0 0 0
Wake kswapd requests 187649 132338 191695 267701
Kswapd wakeups 3 1 4 1
Kswapd pages scanned 4599269 4454162 4296815 3891906
Kswapd pages reclaimed 2295947 2428434 2399818 2319706
Kswapd reclaim write file async I/O 1 0 1 1
Kswapd reclaim write anon async I/O 59 187 41 222
Kswapd reclaim write file sync I/O 0 0 0 0
Kswapd reclaim write anon sync I/O 0 0 0 0
Time stalled direct reclaim (seconds) 4.34 2.52 6.63 2.96
Time kswapd awake (seconds) 11.15 10.25 11.01 10.19
Total pages scanned 4905237 4734564 4897640 4849839
Total pages reclaimed 2339450 2447439 2430145 2436897
%age total pages scanned/reclaimed 47.69% 51.69% 49.62% 50.25%
%age total pages scanned/written 0.00% 0.00% 0.00% 0.00%
%age file pages scanned/written 0.00% 0.00% 0.00% 0.00%
Percentage Time Spent Direct Reclaim 29.23% 19.02% 38.48% 20.25%
Percentage Time kswapd Awake 78.58% 78.85% 76.83% 79.86%
What is interesting here for nocongest in particular is that while direct
reclaim scans more pages, the overall number of pages scanned remains the
same and the ratio of pages scanned to pages reclaimed is more or less the
same. In other words, while we are sleeping less, reclaim is not doing
more work and as direct reclaim and kswapd is awake for less time, it
would appear to be doing less work.
FTrace Reclaim Statistics: congestion_wait
Direct number congest waited 87 196 64 0
Direct time congest waited 4604ms 4732ms 5420ms 0ms
Direct full congest waited 72 145 53 0
Direct number conditional waited 0 0 324 1315
Direct time conditional waited 0ms 0ms 0ms 0ms
Direct full conditional waited 0 0 0 0
KSwapd number congest waited 20 10 15 7
KSwapd time congest waited 1264ms 536ms 884ms 284ms
KSwapd full congest waited 10 4 6 2
KSwapd number conditional waited 0 0 0 0
KSwapd time conditional waited 0ms 0ms 0ms 0ms
KSwapd full conditional waited 0 0 0 0
The vanilla kernel spent 8 seconds asleep in direct reclaim and no time at
all asleep with the patches.
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 10.51 10.73 10.6 11.66
Total Elapsed Time (seconds) 14.19 13.00 14.33 12.76
Overall, the tests completed faster. It is interesting to note that backing off further
when a zone is congested and not just a BDI was more efficient overall.
PPC64 micro-mapped-file-stream
pgalloc_dma 3024660.00 ( 0.00%) 3027185.00 ( 0.08%) 3025845.00 ( 0.04%) 3026281.00 ( 0.05%)
pgalloc_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
pgsteal_dma 2508073.00 ( 0.00%) 2565351.00 ( 2.23%) 2463577.00 ( -1.81%) 2532263.00 ( 0.96%)
pgsteal_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
pgscan_kswapd_dma 4601307.00 ( 0.00%) 4128076.00 ( -11.46%) 3912317.00 ( -17.61%) 3377165.00 ( -36.25%)
pgscan_kswapd_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
pgscan_direct_dma 629825.00 ( 0.00%) 971622.00 ( 35.18%) 1063938.00 ( 40.80%) 1711935.00 ( 63.21%)
pgscan_direct_normal 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%) 0.00 ( 0.00%)
pageoutrun 27776.00 ( 0.00%) 20458.00 ( -35.77%) 18763.00 ( -48.04%) 18157.00 ( -52.98%)
allocstall 977.00 ( 0.00%) 2751.00 ( 64.49%) 2098.00 ( 53.43%) 5136.00 ( 80.98%)
Similar trends to x86-64. allocstalls are up but it's not necessarily bad.
FTrace Reclaim Statistics: vmscan
Direct reclaims 977 2709 2098 5136
Direct reclaim pages scanned 629825 963814 1063938 1711935
Direct reclaim pages reclaimed 75550 242538 150904 387647
Direct reclaim write file async I/O 0 0 0 2
Direct reclaim write anon async I/O 0 10 0 4
Direct reclaim write file sync I/O 0 0 0 0
Direct reclaim write anon sync I/O 0 0 0 0
Wake kswapd requests 392119 1201712 571935 571921
Kswapd wakeups 3 2 3 3
Kswapd pages scanned 4601307 4128076 3912317 3377165
Kswapd pages reclaimed 2432523 2318797 2312673 2144616
Kswapd reclaim write file async I/O 20 1 1 1
Kswapd reclaim write anon async I/O 57 132 11 121
Kswapd reclaim write file sync I/O 0 0 0 0
Kswapd reclaim write anon sync I/O 0 0 0 0
Time stalled direct reclaim (seconds) 6.19 7.30 13.04 10.88
Time kswapd awake (seconds) 21.73 26.51 25.55 23.90
Total pages scanned 5231132 5091890 4976255 5089100
Total pages reclaimed 2508073 2561335 2463577 2532263
%age total pages scanned/reclaimed 47.95% 50.30% 49.51% 49.76%
%age total pages scanned/written 0.00% 0.00% 0.00% 0.00%
%age file pages scanned/written 0.00% 0.00% 0.00% 0.00%
Percentage Time Spent Direct Reclaim 18.89% 20.65% 32.65% 27.65%
Percentage Time kswapd Awake 72.39% 80.68% 78.21% 77.40%
Again, a similar trend that the congestion_wait changes mean that direct
reclaim scans more pages but the overall number of pages scanned while
slightly reduced, are very similar. The ratio of scanning/reclaimed
remains roughly similar. The downside is that kswapd and direct reclaim
was awake longer and for a larger percentage of the overall workload.
It's possible there were big differences in the amount of time spent
reclaiming slab pages between the different kernels which is plausible
considering that the micro tests runs after fsmark and sysbench.
Trace Reclaim Statistics: congestion_wait
Direct number congest waited 845 1312 104 0
Direct time congest waited 19416ms 26560ms 7544ms 0ms
Direct full congest waited 745 1105 72 0
Direct number conditional waited 0 0 1322 2935
Direct time conditional waited 0ms 0ms 12ms 312ms
Direct full conditional waited 0 0 0 3
KSwapd number congest waited 39 102 75 63
KSwapd time congest waited 2484ms 6760ms 5756ms 3716ms
KSwapd full congest waited 20 48 46 25
KSwapd number conditional waited 0 0 0 0
KSwapd time conditional waited 0ms 0ms 0ms 0ms
KSwapd full conditional waited 0 0 0 0
The vanilla kernel spent 20 seconds asleep in direct reclaim and only
312ms asleep with the patches. The time kswapd spent congest waited was
also reduced by a large factor.
MMTests Statistics: duration
ser/Sys Time Running Test (seconds) 26.58 28.05 26.9 28.47
Total Elapsed Time (seconds) 30.02 32.86 32.67 30.88
With all patches applies, the completion times are very similar.
X86-64 STRESS-HIGHALLOC
traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4
Pass 1 82.00 ( 0.00%) 84.00 ( 2.00%) 85.00 ( 3.00%) 85.00 ( 3.00%)
Pass 2 90.00 ( 0.00%) 87.00 (-3.00%) 88.00 (-2.00%) 89.00 (-1.00%)
At Rest 92.00 ( 0.00%) 90.00 (-2.00%) 90.00 (-2.00%) 91.00 (-1.00%)
Success figures across the board are broadly similar.
traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4
Direct reclaims 1045 944 886 887
Direct reclaim pages scanned 135091 119604 109382 101019
Direct reclaim pages reclaimed 88599 47535 47863 46671
Direct reclaim write file async I/O 494 283 465 280
Direct reclaim write anon async I/O 29357 13710 16656 13462
Direct reclaim write file sync I/O 154 2 2 3
Direct reclaim write anon sync I/O 14594 571 509 561
Wake kswapd requests 7491 933 872 892
Kswapd wakeups 814 778 731 780
Kswapd pages scanned 7290822 15341158 11916436 13703442
Kswapd pages reclaimed 3587336 3142496 3094392 3187151
Kswapd reclaim write file async I/O 91975 32317 28022 29628
Kswapd reclaim write anon async I/O 1992022 789307 829745 849769
Kswapd reclaim write file sync I/O 0 0 0 0
Kswapd reclaim write anon sync I/O 0 0 0 0
Time stalled direct reclaim (seconds) 4588.93 2467.16 2495.41 2547.07
Time kswapd awake (seconds) 2497.66 1020.16 1098.06 1176.82
Total pages scanned 7425913 15460762 12025818 13804461
Total pages reclaimed 3675935 3190031 3142255 3233822
%age total pages scanned/reclaimed 49.50% 20.63% 26.13% 23.43%
%age total pages scanned/written 28.66% 5.41% 7.28% 6.47%
%age file pages scanned/written 1.25% 0.21% 0.24% 0.22%
Percentage Time Spent Direct Reclaim 57.33% 42.15% 42.41% 42.99%
Percentage Time kswapd Awake 43.56% 27.87% 29.76% 31.25%
Scanned/reclaimed ratios again look good with big improvements in
efficiency. The Scanned/written ratios also look much improved. With a
better scanned/written ration, there is an expectation that IO would be
more efficient and indeed, the time spent in direct reclaim is much
reduced by the full series and kswapd spends a little less time awake.
Overall, indications here are that allocations were happening much faster
and this can be seen with a graph of the latency figures as the
allocations were taking place
http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-hydra-mean.ps
FTrace Reclaim Statistics: congestion_wait
Direct number congest waited 1333 204 169 4
Direct time congest waited 78896ms 8288ms 7260ms 200ms
Direct full congest waited 756 92 69 2
Direct number conditional waited 0 0 26 186
Direct time conditional waited 0ms 0ms 0ms 2504ms
Direct full conditional waited 0 0 0 25
KSwapd number congest waited 4 395 227 282
KSwapd time congest waited 384ms 25136ms 10508ms 18380ms
KSwapd full congest waited 3 232 98 176
KSwapd number conditional waited 0 0 0 0
KSwapd time conditional waited 0ms 0ms 0ms 0ms
KSwapd full conditional waited 0 0 0 0
KSwapd full conditional waited 318 0 312 9
Overall, the time spent speeping is reduced. kswapd is still hitting
congestion_wait() but that is because there are callers remaining where it
wasn't clear in advance if they should be changed to wait_iff_congested()
or not. Overall the sleep imes are reduced though - from 79ish seconds to
about 19.
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 3415.43 3386.65 3388.39 3377.5
Total Elapsed Time (seconds) 5733.48 3660.33 3689.41 3765.39
With the full series, the time to complete the tests are reduced by 30%
PPC64 STRESS-HIGHALLOC
traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4
Pass 1 17.00 ( 0.00%) 34.00 (17.00%) 38.00 (21.00%) 43.00 (26.00%)
Pass 2 25.00 ( 0.00%) 37.00 (12.00%) 42.00 (17.00%) 46.00 (21.00%)
At Rest 49.00 ( 0.00%) 43.00 (-6.00%) 45.00 (-4.00%) 51.00 ( 2.00%)
Success rates there are *way* up particularly considering that the 16MB
huge pages on PPC64 mean that it's always much harder to allocate them.
FTrace Reclaim Statistics: vmscan
stress-highalloc stress-highalloc stress-highalloc stress-highalloc
traceonly-v2r2 lowlumpy-v2r3 waitcongest-v2r3waitwriteback-v2r4
Direct reclaims 499 505 564 509
Direct reclaim pages scanned 223478 41898 51818 45605
Direct reclaim pages reclaimed 137730 21148 27161 23455
Direct reclaim write file async I/O 399 136 162 136
Direct reclaim write anon async I/O 46977 2865 4686 3998
Direct reclaim write file sync I/O 29 0 1 3
Direct reclaim write anon sync I/O 31023 159 237 239
Wake kswapd requests 420 351 360 326
Kswapd wakeups 185 294 249 277
Kswapd pages scanned 15703488 16392500 17821724 17598737
Kswapd pages reclaimed 5808466 2908858 3139386 3145435
Kswapd reclaim write file async I/O 159938 18400 18717 13473
Kswapd reclaim write anon async I/O 3467554 228957 322799 234278
Kswapd reclaim write file sync I/O 0 0 0 0
Kswapd reclaim write anon sync I/O 0 0 0 0
Time stalled direct reclaim (seconds) 9665.35 1707.81 2374.32 1871.23
Time kswapd awake (seconds) 9401.21 1367.86 1951.75 1328.88
Total pages scanned 15926966 16434398 17873542 17644342
Total pages reclaimed 5946196 2930006 3166547 3168890
%age total pages scanned/reclaimed 37.33% 17.83% 17.72% 17.96%
%age total pages scanned/written 23.27% 1.52% 1.94% 1.43%
%age file pages scanned/written 1.01% 0.11% 0.11% 0.08%
Percentage Time Spent Direct Reclaim 44.55% 35.10% 41.42% 36.91%
Percentage Time kswapd Awake 86.71% 43.58% 52.67% 41.14%
While the scanning rates are slightly up, the scanned/reclaimed and
scanned/written figures are much improved. The time spent in direct
reclaim and with kswapd are massively reduced, mostly by the lowlumpy
patches.
FTrace Reclaim Statistics: congestion_wait
Direct number congest waited 725 303 126 3
Direct time congest waited 45524ms 9180ms 5936ms 300ms
Direct full congest waited 487 190 52 3
Direct number conditional waited 0 0 200 301
Direct time conditional waited 0ms 0ms 0ms 1904ms
Direct full conditional waited 0 0 0 19
KSwapd number congest waited 0 2 23 4
KSwapd time congest waited 0ms 200ms 420ms 404ms
KSwapd full congest waited 0 2 2 4
KSwapd number conditional waited 0 0 0 0
KSwapd time conditional waited 0ms 0ms 0ms 0ms
KSwapd full conditional waited 0 0 0 0
Not as dramatic a story here but the time spent asleep is reduced and we
can still see what wait_iff_congested is going to sleep when necessary.
MMTests Statistics: duration
User/Sys Time Running Test (seconds) 12028.09 3157.17 3357.79 3199.16
Total Elapsed Time (seconds) 10842.07 3138.72 3705.54 3229.85
The time to complete this test goes way down. With the full series, we
are allocating over twice the number of huge pages in 30% of the time and
there is a corresponding impact on the allocation latency graph available
at.
http://www.csn.ul.ie/~mel/postings/vmscanreduce-20101509/highalloc-interlatency-powyah-mean.ps
This patch:
Add a trace event for shrink_inactive_list() and updates the sample
postprocessing script appropriately. It can be used to determine how many
pages were reclaimed and for non-lumpy reclaim where exactly the pages
were reclaimed from.
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-26 21:21:40 +00:00
|
|
|
|
mm/vmscan: wake up flushers for legacy cgroups too
Commit 726d061fbd36 ("mm: vmscan: kick flushers when we encounter dirty
pages on the LRU") added flusher invocation to shrink_inactive_list()
when many dirty pages on the LRU are encountered.
However, shrink_inactive_list() doesn't wake up flushers for legacy
cgroup reclaim, so the next commit bbef938429f5 ("mm: vmscan: remove old
flusher wakeup from direct reclaim path") removed the only source of
flusher's wake up in legacy mem cgroup reclaim path.
This leads to premature OOM if there is too many dirty pages in cgroup:
# mkdir /sys/fs/cgroup/memory/test
# echo $$ > /sys/fs/cgroup/memory/test/tasks
# echo 50M > /sys/fs/cgroup/memory/test/memory.limit_in_bytes
# dd if=/dev/zero of=tmp_file bs=1M count=100
Killed
dd invoked oom-killer: gfp_mask=0x14000c0(GFP_KERNEL), nodemask=(null), order=0, oom_score_adj=0
Call Trace:
dump_stack+0x46/0x65
dump_header+0x6b/0x2ac
oom_kill_process+0x21c/0x4a0
out_of_memory+0x2a5/0x4b0
mem_cgroup_out_of_memory+0x3b/0x60
mem_cgroup_oom_synchronize+0x2ed/0x330
pagefault_out_of_memory+0x24/0x54
__do_page_fault+0x521/0x540
page_fault+0x45/0x50
Task in /test killed as a result of limit of /test
memory: usage 51200kB, limit 51200kB, failcnt 73
memory+swap: usage 51200kB, limit 9007199254740988kB, failcnt 0
kmem: usage 296kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /test: cache:49632KB rss:1056KB rss_huge:0KB shmem:0KB
mapped_file:0KB dirty:49500KB writeback:0KB swap:0KB inactive_anon:0KB
active_anon:1168KB inactive_file:24760KB active_file:24960KB unevictable:0KB
Memory cgroup out of memory: Kill process 3861 (bash) score 88 or sacrifice child
Killed process 3876 (dd) total-vm:8484kB, anon-rss:1052kB, file-rss:1720kB, shmem-rss:0kB
oom_reaper: reaped process 3876 (dd), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
Wake up flushers in legacy cgroup reclaim too.
Link: http://lkml.kernel.org/r/20180315164553.17856-1-aryabinin@virtuozzo.com
Fixes: bbef938429f5 ("mm: vmscan: remove old flusher wakeup from direct reclaim path")
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-03-22 23:17:42 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* If dirty folios are scanned that are not queued for IO, it
|
mm/vmscan: wake up flushers for legacy cgroups too
Commit 726d061fbd36 ("mm: vmscan: kick flushers when we encounter dirty
pages on the LRU") added flusher invocation to shrink_inactive_list()
when many dirty pages on the LRU are encountered.
However, shrink_inactive_list() doesn't wake up flushers for legacy
cgroup reclaim, so the next commit bbef938429f5 ("mm: vmscan: remove old
flusher wakeup from direct reclaim path") removed the only source of
flusher's wake up in legacy mem cgroup reclaim path.
This leads to premature OOM if there is too many dirty pages in cgroup:
# mkdir /sys/fs/cgroup/memory/test
# echo $$ > /sys/fs/cgroup/memory/test/tasks
# echo 50M > /sys/fs/cgroup/memory/test/memory.limit_in_bytes
# dd if=/dev/zero of=tmp_file bs=1M count=100
Killed
dd invoked oom-killer: gfp_mask=0x14000c0(GFP_KERNEL), nodemask=(null), order=0, oom_score_adj=0
Call Trace:
dump_stack+0x46/0x65
dump_header+0x6b/0x2ac
oom_kill_process+0x21c/0x4a0
out_of_memory+0x2a5/0x4b0
mem_cgroup_out_of_memory+0x3b/0x60
mem_cgroup_oom_synchronize+0x2ed/0x330
pagefault_out_of_memory+0x24/0x54
__do_page_fault+0x521/0x540
page_fault+0x45/0x50
Task in /test killed as a result of limit of /test
memory: usage 51200kB, limit 51200kB, failcnt 73
memory+swap: usage 51200kB, limit 9007199254740988kB, failcnt 0
kmem: usage 296kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /test: cache:49632KB rss:1056KB rss_huge:0KB shmem:0KB
mapped_file:0KB dirty:49500KB writeback:0KB swap:0KB inactive_anon:0KB
active_anon:1168KB inactive_file:24760KB active_file:24960KB unevictable:0KB
Memory cgroup out of memory: Kill process 3861 (bash) score 88 or sacrifice child
Killed process 3876 (dd) total-vm:8484kB, anon-rss:1052kB, file-rss:1720kB, shmem-rss:0kB
oom_reaper: reaped process 3876 (dd), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
Wake up flushers in legacy cgroup reclaim too.
Link: http://lkml.kernel.org/r/20180315164553.17856-1-aryabinin@virtuozzo.com
Fixes: bbef938429f5 ("mm: vmscan: remove old flusher wakeup from direct reclaim path")
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-03-22 23:17:42 +00:00
|
|
|
* implies that flushers are not doing their job. This can
|
2022-09-02 19:45:57 +00:00
|
|
|
* happen when memory pressure pushes dirty folios to the end of
|
mm/vmscan: wake up flushers for legacy cgroups too
Commit 726d061fbd36 ("mm: vmscan: kick flushers when we encounter dirty
pages on the LRU") added flusher invocation to shrink_inactive_list()
when many dirty pages on the LRU are encountered.
However, shrink_inactive_list() doesn't wake up flushers for legacy
cgroup reclaim, so the next commit bbef938429f5 ("mm: vmscan: remove old
flusher wakeup from direct reclaim path") removed the only source of
flusher's wake up in legacy mem cgroup reclaim path.
This leads to premature OOM if there is too many dirty pages in cgroup:
# mkdir /sys/fs/cgroup/memory/test
# echo $$ > /sys/fs/cgroup/memory/test/tasks
# echo 50M > /sys/fs/cgroup/memory/test/memory.limit_in_bytes
# dd if=/dev/zero of=tmp_file bs=1M count=100
Killed
dd invoked oom-killer: gfp_mask=0x14000c0(GFP_KERNEL), nodemask=(null), order=0, oom_score_adj=0
Call Trace:
dump_stack+0x46/0x65
dump_header+0x6b/0x2ac
oom_kill_process+0x21c/0x4a0
out_of_memory+0x2a5/0x4b0
mem_cgroup_out_of_memory+0x3b/0x60
mem_cgroup_oom_synchronize+0x2ed/0x330
pagefault_out_of_memory+0x24/0x54
__do_page_fault+0x521/0x540
page_fault+0x45/0x50
Task in /test killed as a result of limit of /test
memory: usage 51200kB, limit 51200kB, failcnt 73
memory+swap: usage 51200kB, limit 9007199254740988kB, failcnt 0
kmem: usage 296kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /test: cache:49632KB rss:1056KB rss_huge:0KB shmem:0KB
mapped_file:0KB dirty:49500KB writeback:0KB swap:0KB inactive_anon:0KB
active_anon:1168KB inactive_file:24760KB active_file:24960KB unevictable:0KB
Memory cgroup out of memory: Kill process 3861 (bash) score 88 or sacrifice child
Killed process 3876 (dd) total-vm:8484kB, anon-rss:1052kB, file-rss:1720kB, shmem-rss:0kB
oom_reaper: reaped process 3876 (dd), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
Wake up flushers in legacy cgroup reclaim too.
Link: http://lkml.kernel.org/r/20180315164553.17856-1-aryabinin@virtuozzo.com
Fixes: bbef938429f5 ("mm: vmscan: remove old flusher wakeup from direct reclaim path")
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-03-22 23:17:42 +00:00
|
|
|
* the LRU before the dirty limits are breached and the dirty
|
|
|
|
* data has expired. It can also happen when the proportion of
|
2022-09-02 19:45:57 +00:00
|
|
|
* dirty folios grows not through writes but through memory
|
mm/vmscan: wake up flushers for legacy cgroups too
Commit 726d061fbd36 ("mm: vmscan: kick flushers when we encounter dirty
pages on the LRU") added flusher invocation to shrink_inactive_list()
when many dirty pages on the LRU are encountered.
However, shrink_inactive_list() doesn't wake up flushers for legacy
cgroup reclaim, so the next commit bbef938429f5 ("mm: vmscan: remove old
flusher wakeup from direct reclaim path") removed the only source of
flusher's wake up in legacy mem cgroup reclaim path.
This leads to premature OOM if there is too many dirty pages in cgroup:
# mkdir /sys/fs/cgroup/memory/test
# echo $$ > /sys/fs/cgroup/memory/test/tasks
# echo 50M > /sys/fs/cgroup/memory/test/memory.limit_in_bytes
# dd if=/dev/zero of=tmp_file bs=1M count=100
Killed
dd invoked oom-killer: gfp_mask=0x14000c0(GFP_KERNEL), nodemask=(null), order=0, oom_score_adj=0
Call Trace:
dump_stack+0x46/0x65
dump_header+0x6b/0x2ac
oom_kill_process+0x21c/0x4a0
out_of_memory+0x2a5/0x4b0
mem_cgroup_out_of_memory+0x3b/0x60
mem_cgroup_oom_synchronize+0x2ed/0x330
pagefault_out_of_memory+0x24/0x54
__do_page_fault+0x521/0x540
page_fault+0x45/0x50
Task in /test killed as a result of limit of /test
memory: usage 51200kB, limit 51200kB, failcnt 73
memory+swap: usage 51200kB, limit 9007199254740988kB, failcnt 0
kmem: usage 296kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /test: cache:49632KB rss:1056KB rss_huge:0KB shmem:0KB
mapped_file:0KB dirty:49500KB writeback:0KB swap:0KB inactive_anon:0KB
active_anon:1168KB inactive_file:24760KB active_file:24960KB unevictable:0KB
Memory cgroup out of memory: Kill process 3861 (bash) score 88 or sacrifice child
Killed process 3876 (dd) total-vm:8484kB, anon-rss:1052kB, file-rss:1720kB, shmem-rss:0kB
oom_reaper: reaped process 3876 (dd), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
Wake up flushers in legacy cgroup reclaim too.
Link: http://lkml.kernel.org/r/20180315164553.17856-1-aryabinin@virtuozzo.com
Fixes: bbef938429f5 ("mm: vmscan: remove old flusher wakeup from direct reclaim path")
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-03-22 23:17:42 +00:00
|
|
|
* pressure reclaiming all the clean cache. And in some cases,
|
|
|
|
* the flushers simply cannot keep up with the allocation
|
|
|
|
* rate. Nudge the flusher threads in case they are asleep.
|
|
|
|
*/
|
2022-11-18 07:06:03 +00:00
|
|
|
if (stat.nr_unqueued_dirty == nr_taken) {
|
mm/vmscan: wake up flushers for legacy cgroups too
Commit 726d061fbd36 ("mm: vmscan: kick flushers when we encounter dirty
pages on the LRU") added flusher invocation to shrink_inactive_list()
when many dirty pages on the LRU are encountered.
However, shrink_inactive_list() doesn't wake up flushers for legacy
cgroup reclaim, so the next commit bbef938429f5 ("mm: vmscan: remove old
flusher wakeup from direct reclaim path") removed the only source of
flusher's wake up in legacy mem cgroup reclaim path.
This leads to premature OOM if there is too many dirty pages in cgroup:
# mkdir /sys/fs/cgroup/memory/test
# echo $$ > /sys/fs/cgroup/memory/test/tasks
# echo 50M > /sys/fs/cgroup/memory/test/memory.limit_in_bytes
# dd if=/dev/zero of=tmp_file bs=1M count=100
Killed
dd invoked oom-killer: gfp_mask=0x14000c0(GFP_KERNEL), nodemask=(null), order=0, oom_score_adj=0
Call Trace:
dump_stack+0x46/0x65
dump_header+0x6b/0x2ac
oom_kill_process+0x21c/0x4a0
out_of_memory+0x2a5/0x4b0
mem_cgroup_out_of_memory+0x3b/0x60
mem_cgroup_oom_synchronize+0x2ed/0x330
pagefault_out_of_memory+0x24/0x54
__do_page_fault+0x521/0x540
page_fault+0x45/0x50
Task in /test killed as a result of limit of /test
memory: usage 51200kB, limit 51200kB, failcnt 73
memory+swap: usage 51200kB, limit 9007199254740988kB, failcnt 0
kmem: usage 296kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /test: cache:49632KB rss:1056KB rss_huge:0KB shmem:0KB
mapped_file:0KB dirty:49500KB writeback:0KB swap:0KB inactive_anon:0KB
active_anon:1168KB inactive_file:24760KB active_file:24960KB unevictable:0KB
Memory cgroup out of memory: Kill process 3861 (bash) score 88 or sacrifice child
Killed process 3876 (dd) total-vm:8484kB, anon-rss:1052kB, file-rss:1720kB, shmem-rss:0kB
oom_reaper: reaped process 3876 (dd), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
Wake up flushers in legacy cgroup reclaim too.
Link: http://lkml.kernel.org/r/20180315164553.17856-1-aryabinin@virtuozzo.com
Fixes: bbef938429f5 ("mm: vmscan: remove old flusher wakeup from direct reclaim path")
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-03-22 23:17:42 +00:00
|
|
|
wakeup_flusher_threads(WB_REASON_VMSCAN);
|
2022-11-18 07:06:03 +00:00
|
|
|
/*
|
|
|
|
* For cgroupv1 dirty throttling is achieved by waking up
|
|
|
|
* the kernel flusher here and later waiting on folios
|
|
|
|
* which are in writeback to finish (see shrink_folio_list()).
|
|
|
|
*
|
|
|
|
* Flusher may not be able to issue writeback quickly
|
|
|
|
* enough for cgroupv1 writeback throttling to work
|
|
|
|
* on a large system.
|
|
|
|
*/
|
|
|
|
if (!writeback_throttling_sane(sc))
|
|
|
|
reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
|
|
|
|
}
|
mm/vmscan: wake up flushers for legacy cgroups too
Commit 726d061fbd36 ("mm: vmscan: kick flushers when we encounter dirty
pages on the LRU") added flusher invocation to shrink_inactive_list()
when many dirty pages on the LRU are encountered.
However, shrink_inactive_list() doesn't wake up flushers for legacy
cgroup reclaim, so the next commit bbef938429f5 ("mm: vmscan: remove old
flusher wakeup from direct reclaim path") removed the only source of
flusher's wake up in legacy mem cgroup reclaim path.
This leads to premature OOM if there is too many dirty pages in cgroup:
# mkdir /sys/fs/cgroup/memory/test
# echo $$ > /sys/fs/cgroup/memory/test/tasks
# echo 50M > /sys/fs/cgroup/memory/test/memory.limit_in_bytes
# dd if=/dev/zero of=tmp_file bs=1M count=100
Killed
dd invoked oom-killer: gfp_mask=0x14000c0(GFP_KERNEL), nodemask=(null), order=0, oom_score_adj=0
Call Trace:
dump_stack+0x46/0x65
dump_header+0x6b/0x2ac
oom_kill_process+0x21c/0x4a0
out_of_memory+0x2a5/0x4b0
mem_cgroup_out_of_memory+0x3b/0x60
mem_cgroup_oom_synchronize+0x2ed/0x330
pagefault_out_of_memory+0x24/0x54
__do_page_fault+0x521/0x540
page_fault+0x45/0x50
Task in /test killed as a result of limit of /test
memory: usage 51200kB, limit 51200kB, failcnt 73
memory+swap: usage 51200kB, limit 9007199254740988kB, failcnt 0
kmem: usage 296kB, limit 9007199254740988kB, failcnt 0
Memory cgroup stats for /test: cache:49632KB rss:1056KB rss_huge:0KB shmem:0KB
mapped_file:0KB dirty:49500KB writeback:0KB swap:0KB inactive_anon:0KB
active_anon:1168KB inactive_file:24760KB active_file:24960KB unevictable:0KB
Memory cgroup out of memory: Kill process 3861 (bash) score 88 or sacrifice child
Killed process 3876 (dd) total-vm:8484kB, anon-rss:1052kB, file-rss:1720kB, shmem-rss:0kB
oom_reaper: reaped process 3876 (dd), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
Wake up flushers in legacy cgroup reclaim too.
Link: http://lkml.kernel.org/r/20180315164553.17856-1-aryabinin@virtuozzo.com
Fixes: bbef938429f5 ("mm: vmscan: remove old flusher wakeup from direct reclaim path")
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Tested-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-03-22 23:17:42 +00:00
|
|
|
|
2018-04-10 23:27:59 +00:00
|
|
|
sc->nr.dirty += stat.nr_dirty;
|
|
|
|
sc->nr.congested += stat.nr_congested;
|
|
|
|
sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
|
|
|
|
sc->nr.writeback += stat.nr_writeback;
|
|
|
|
sc->nr.immediate += stat.nr_immediate;
|
|
|
|
sc->nr.taken += nr_taken;
|
|
|
|
if (file)
|
|
|
|
sc->nr.file_taken += nr_taken;
|
2013-07-03 22:02:02 +00:00
|
|
|
|
2016-07-28 22:45:31 +00:00
|
|
|
trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
|
mm, vmscan, tracing: use pointer to reclaim_stat struct in trace event
The trace event trace_mm_vmscan_lru_shrink_inactive() currently has 12
parameters! Seven of them are from the reclaim_stat structure. This
structure is currently local to mm/vmscan.c. By moving it to the global
vmstat.h header, we can also reference it from the vmscan tracepoints.
In moving it, it brings down the overhead of passing so many arguments
to the trace event. In the future, we may limit the number of arguments
that a trace event may pass (ideally just 6, but more realistically it
may be 8).
Before this patch, the code to call the trace event is this:
0f 83 aa fe ff ff jae ffffffff811e6261 <shrink_inactive_list+0x1e1>
48 8b 45 a0 mov -0x60(%rbp),%rax
45 8b 64 24 20 mov 0x20(%r12),%r12d
44 8b 6d d4 mov -0x2c(%rbp),%r13d
8b 4d d0 mov -0x30(%rbp),%ecx
44 8b 75 cc mov -0x34(%rbp),%r14d
44 8b 7d c8 mov -0x38(%rbp),%r15d
48 89 45 90 mov %rax,-0x70(%rbp)
8b 83 b8 fe ff ff mov -0x148(%rbx),%eax
8b 55 c0 mov -0x40(%rbp),%edx
8b 7d c4 mov -0x3c(%rbp),%edi
8b 75 b8 mov -0x48(%rbp),%esi
89 45 80 mov %eax,-0x80(%rbp)
65 ff 05 e4 f7 e2 7e incl %gs:0x7ee2f7e4(%rip) # 15bd0 <__preempt_count>
48 8b 05 75 5b 13 01 mov 0x1135b75(%rip),%rax # ffffffff8231bf68 <__tracepoint_mm_vmscan_lru_shrink_inactive+0x28>
48 85 c0 test %rax,%rax
74 72 je ffffffff811e646a <shrink_inactive_list+0x3ea>
48 89 c3 mov %rax,%rbx
4c 8b 10 mov (%rax),%r10
89 f8 mov %edi,%eax
48 89 85 68 ff ff ff mov %rax,-0x98(%rbp)
89 f0 mov %esi,%eax
48 89 85 60 ff ff ff mov %rax,-0xa0(%rbp)
89 c8 mov %ecx,%eax
48 89 85 78 ff ff ff mov %rax,-0x88(%rbp)
89 d0 mov %edx,%eax
48 89 85 70 ff ff ff mov %rax,-0x90(%rbp)
8b 45 8c mov -0x74(%rbp),%eax
48 8b 7b 08 mov 0x8(%rbx),%rdi
48 83 c3 18 add $0x18,%rbx
50 push %rax
41 54 push %r12
41 55 push %r13
ff b5 78 ff ff ff pushq -0x88(%rbp)
41 56 push %r14
41 57 push %r15
ff b5 70 ff ff ff pushq -0x90(%rbp)
4c 8b 8d 68 ff ff ff mov -0x98(%rbp),%r9
4c 8b 85 60 ff ff ff mov -0xa0(%rbp),%r8
48 8b 4d 98 mov -0x68(%rbp),%rcx
48 8b 55 90 mov -0x70(%rbp),%rdx
8b 75 80 mov -0x80(%rbp),%esi
41 ff d2 callq *%r10
After the patch:
0f 83 a8 fe ff ff jae ffffffff811e626d <shrink_inactive_list+0x1cd>
8b 9b b8 fe ff ff mov -0x148(%rbx),%ebx
45 8b 64 24 20 mov 0x20(%r12),%r12d
4c 8b 6d a0 mov -0x60(%rbp),%r13
65 ff 05 f5 f7 e2 7e incl %gs:0x7ee2f7f5(%rip) # 15bd0 <__preempt_count>
4c 8b 35 86 5b 13 01 mov 0x1135b86(%rip),%r14 # ffffffff8231bf68 <__tracepoint_mm_vmscan_lru_shrink_inactive+0x28>
4d 85 f6 test %r14,%r14
74 2a je ffffffff811e6411 <shrink_inactive_list+0x371>
49 8b 06 mov (%r14),%rax
8b 4d 8c mov -0x74(%rbp),%ecx
49 8b 7e 08 mov 0x8(%r14),%rdi
49 83 c6 18 add $0x18,%r14
4c 89 ea mov %r13,%rdx
45 89 e1 mov %r12d,%r9d
4c 8d 45 b8 lea -0x48(%rbp),%r8
89 de mov %ebx,%esi
51 push %rcx
48 8b 4d 98 mov -0x68(%rbp),%rcx
ff d0 callq *%rax
Link: http://lkml.kernel.org/r/2559d7cb-ec60-1200-2362-04fa34fd02bb@fb.com
Link: http://lkml.kernel.org/r/20180322121003.4177af15@gandalf.local.home
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Reported-by: Alexei Starovoitov <ast@fb.com>
Acked-by: David Rientjes <rientjes@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexei Starovoitov <ast@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-10 23:28:07 +00:00
|
|
|
nr_scanned, nr_reclaimed, &stat, sc->priority, file);
|
2006-03-22 08:08:20 +00:00
|
|
|
return nr_reclaimed;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2020-12-15 22:21:31 +00:00
|
|
|
/*
|
2022-06-17 15:42:47 +00:00
|
|
|
* shrink_active_list() moves folios from the active LRU to the inactive LRU.
|
2020-12-15 22:21:31 +00:00
|
|
|
*
|
2022-06-17 15:42:47 +00:00
|
|
|
* We move them the other way if the folio is referenced by one or more
|
2020-12-15 22:21:31 +00:00
|
|
|
* processes.
|
|
|
|
*
|
2022-06-17 15:42:47 +00:00
|
|
|
* If the folios are mostly unmapped, the processing is fast and it is
|
2020-12-15 22:21:31 +00:00
|
|
|
* appropriate to hold lru_lock across the whole operation. But if
|
2022-06-17 15:42:47 +00:00
|
|
|
* the folios are mapped, the processing is slow (folio_referenced()), so
|
|
|
|
* we should drop lru_lock around each folio. It's impossible to balance
|
|
|
|
* this, so instead we remove the folios from the LRU while processing them.
|
|
|
|
* It is safe to rely on the active flag against the non-LRU folios in here
|
|
|
|
* because nobody will play with that bit on a non-LRU folio.
|
2020-12-15 22:21:31 +00:00
|
|
|
*
|
2022-06-17 15:42:47 +00:00
|
|
|
* The downside is that we have to touch folio->_refcount against each folio.
|
|
|
|
* But we had to alter folio->flags anyway.
|
2020-12-15 22:21:31 +00:00
|
|
|
*/
|
2012-01-13 01:20:06 +00:00
|
|
|
static void shrink_active_list(unsigned long nr_to_scan,
|
2012-05-29 22:07:01 +00:00
|
|
|
struct lruvec *lruvec,
|
2012-01-13 01:17:52 +00:00
|
|
|
struct scan_control *sc,
|
2012-05-29 22:06:57 +00:00
|
|
|
enum lru_list lru)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2009-09-22 00:01:35 +00:00
|
|
|
unsigned long nr_taken;
|
2012-01-13 01:20:06 +00:00
|
|
|
unsigned long nr_scanned;
|
2009-06-16 22:33:05 +00:00
|
|
|
unsigned long vm_flags;
|
2022-06-17 15:42:47 +00:00
|
|
|
LIST_HEAD(l_hold); /* The folios which were snipped off */
|
vmscan: make mapped executable pages the first class citizen
Protect referenced PROT_EXEC mapped pages from being deactivated.
PROT_EXEC(or its internal presentation VM_EXEC) pages normally belong to some
currently running executables and their linked libraries, they shall really be
cached aggressively to provide good user experiences.
Thanks to Johannes Weiner for the advice to reuse the VMA walk in
page_referenced() to get the PROT_EXEC bit.
[more details]
( The consequences of this patch will have to be discussed together with
Rik van Riel's recent patch "vmscan: evict use-once pages first". )
( Some of the good points and insights are taken into this changelog.
Thanks to all the involved people for the great LKML discussions. )
the problem
===========
For a typical desktop, the most precious working set is composed of
*actively accessed*
(1) memory mapped executables
(2) and their anonymous pages
(3) and other files
(4) and the dcache/icache/.. slabs
while the least important data are
(5) infrequently used or use-once files
For a typical desktop, one major problem is busty and large amount of (5)
use-once files flushing out the working set.
Inside the working set, (4) dcache/icache have already been too sticky ;-)
So we only have to care (2) anonymous and (1)(3) file pages.
anonymous pages
===============
Anonymous pages are effectively immune to the streaming IO attack, because we
now have separate file/anon LRU lists. When the use-once files crowd into the
file LRU, the list's "quality" is significantly lowered. Therefore the scan
balance policy in get_scan_ratio() will choose to scan the (low quality) file
LRU much more frequently than the anon LRU.
file pages
==========
Rik proposed to *not* scan the active file LRU when the inactive list grows
larger than active list. This guarantees that when there are use-once streaming
IO, and the working set is not too large(so that active_size < inactive_size),
the active file LRU will *not* be scanned at all. So the not-too-large working
set can be well protected.
But there are also situations where the file working set is a bit large so that
(active_size >= inactive_size), or the streaming IOs are not purely use-once.
In these cases, the active list will be scanned slowly. Because the current
shrink_active_list() policy is to deactivate active pages regardless of their
referenced bits. The deactivated pages become susceptible to the streaming IO
attack: the inactive list could be scanned fast (500MB / 50MBps = 10s) so that
the deactivated pages don't have enough time to get re-referenced. Because a
user tend to switch between windows in intervals from seconds to minutes.
This patch holds mapped executable pages in the active list as long as they
are referenced during each full scan of the active list. Because the active
list is normally scanned much slower, they get longer grace time (eg. 100s)
for further references, which better matches the pace of user operations.
Therefore this patch greatly prolongs the in-cache time of executable code,
when there are moderate memory pressures.
before patch: guaranteed to be cached if reference intervals < I
after patch: guaranteed to be cached if reference intervals < I+A
(except when randomly reclaimed by the lumpy reclaim)
where
A = time to fully scan the active file LRU
I = time to fully scan the inactive file LRU
Note that normally A >> I.
side effects
============
This patch is safe in general, it restores the pre-2.6.28 mmap() behavior
but in a much smaller and well targeted scope.
One may worry about some one to abuse the PROT_EXEC heuristic. But as
Andrew Morton stated, there are other tricks to getting that sort of boost.
Another concern is the PROT_EXEC mapped pages growing large in rare cases,
and therefore hurting reclaim efficiency. But a sane application targeted for
large audience will never use PROT_EXEC for data mappings. If some home made
application tries to abuse that bit, it shall be aware of the consequences.
If it is abused to scale of 2/3 total memory, it gains nothing but overheads.
benchmarks
==========
1) memory tight desktop
1.1) brief summary
- clock time and major faults are reduced by 50%;
- pswpin numbers are reduced to ~1/3.
That means X desktop responsiveness is doubled under high memory/swap pressure.
1.2) test scenario
- nfsroot gnome desktop with 512M physical memory
- run some programs, and switch between the existing windows
after starting each new program.
1.3) progress timing (seconds)
before after programs
0.02 0.02 N xeyes
0.75 0.76 N firefox
2.02 1.88 N nautilus
3.36 3.17 N nautilus --browser
5.26 4.89 N gthumb
7.12 6.47 N gedit
9.22 8.16 N xpdf /usr/share/doc/shared-mime-info/shared-mime-info-spec.pdf
13.58 12.55 N xterm
15.87 14.57 N mlterm
18.63 17.06 N gnome-terminal
21.16 18.90 N urxvt
26.24 23.48 N gnome-system-monitor
28.72 26.52 N gnome-help
32.15 29.65 N gnome-dictionary
39.66 36.12 N /usr/games/sol
43.16 39.27 N /usr/games/gnometris
48.65 42.56 N /usr/games/gnect
53.31 47.03 N /usr/games/gtali
58.60 52.05 N /usr/games/iagno
65.77 55.42 N /usr/games/gnotravex
70.76 61.47 N /usr/games/mahjongg
76.15 67.11 N /usr/games/gnome-sudoku
86.32 75.15 N /usr/games/glines
92.21 79.70 N /usr/games/glchess
103.79 88.48 N /usr/games/gnomine
113.84 96.51 N /usr/games/gnotski
124.40 102.19 N /usr/games/gnibbles
137.41 114.93 N /usr/games/gnobots2
155.53 125.02 N /usr/games/blackjack
179.85 135.11 N /usr/games/same-gnome
224.49 154.50 N /usr/bin/gnome-window-properties
248.44 162.09 N /usr/bin/gnome-default-applications-properties
282.62 173.29 N /usr/bin/gnome-at-properties
323.72 188.21 N /usr/bin/gnome-typing-monitor
363.99 199.93 N /usr/bin/gnome-at-visual
394.21 206.95 N /usr/bin/gnome-sound-properties
435.14 224.49 N /usr/bin/gnome-at-mobility
463.05 234.11 N /usr/bin/gnome-keybinding-properties
503.75 248.59 N /usr/bin/gnome-about-me
554.00 276.27 N /usr/bin/gnome-display-properties
615.48 304.39 N /usr/bin/gnome-network-preferences
693.03 342.01 N /usr/bin/gnome-mouse-properties
759.90 388.58 N /usr/bin/gnome-appearance-properties
937.90 508.47 N /usr/bin/gnome-control-center
1109.75 587.57 N /usr/bin/gnome-keyboard-properties
1399.05 758.16 N : oocalc
1524.64 830.03 N : oodraw
1684.31 900.03 N : ooimpress
1874.04 993.91 N : oomath
2115.12 1081.89 N : ooweb
2369.02 1161.99 N : oowriter
Note that the last ": oo*" commands are actually commented out.
1.4) vmstat numbers (some relevant ones are marked with *)
before after
nr_free_pages 1293 3898
nr_inactive_anon 59956 53460
nr_active_anon 26815 30026
nr_inactive_file 2657 3218
nr_active_file 2019 2806
nr_unevictable 4 4
nr_mlock 4 4
nr_anon_pages 26706 27859
*nr_mapped 3542 4469
nr_file_pages 72232 67681
nr_dirty 1 0
nr_writeback 123 19
nr_slab_reclaimable 3375 3534
nr_slab_unreclaimable 11405 10665
nr_page_table_pages 8106 7864
nr_unstable 0 0
nr_bounce 0 0
*nr_vmscan_write 394776 230839
nr_writeback_temp 0 0
numa_hit 6843353 3318676
numa_miss 0 0
numa_foreign 0 0
numa_interleave 1719 1719
numa_local 6843353 3318676
numa_other 0 0
*pgpgin 5954683 2057175
*pgpgout 1578276 922744
*pswpin 1486615 512238
*pswpout 394568 230685
pgalloc_dma 277432 56602
pgalloc_dma32 6769477 3310348
pgalloc_normal 0 0
pgalloc_movable 0 0
pgfree 7048396 3371118
pgactivate 2036343 1471492
pgdeactivate 2189691 1612829
pgfault 3702176 3100702
*pgmajfault 452116 201343
pgrefill_dma 12185 7127
pgrefill_dma32 334384 653703
pgrefill_normal 0 0
pgrefill_movable 0 0
pgsteal_dma 74214 22179
pgsteal_dma32 3334164 1638029
pgsteal_normal 0 0
pgsteal_movable 0 0
pgscan_kswapd_dma 1081421 1216199
pgscan_kswapd_dma32 58979118 46002810
pgscan_kswapd_normal 0 0
pgscan_kswapd_movable 0 0
pgscan_direct_dma 2015438 1086109
pgscan_direct_dma32 55787823 36101597
pgscan_direct_normal 0 0
pgscan_direct_movable 0 0
pginodesteal 3461 7281
slabs_scanned 564864 527616
kswapd_steal 2889797 1448082
kswapd_inodesteal 14827 14835
pageoutrun 43459 21562
allocstall 9653 4032
pgrotated 384216 228631
1.5) free numbers at the end of the tests
before patch:
total used free shared buffers cached
Mem: 474 467 7 0 0 236
-/+ buffers/cache: 230 243
Swap: 1023 418 605
after patch:
total used free shared buffers cached
Mem: 474 457 16 0 0 236
-/+ buffers/cache: 221 253
Swap: 1023 404 619
2) memory flushing in a file server
2.1) brief summary
The number of major faults from 50 to 3 during 10% cache hot reads.
That means this patch successfully stops major faults when the active file
list is slowly scanned when there are partially cache hot streaming IO.
2.2) test scenario
Do 100000 pread(size=110 pages, offset=(i*100) pages), where 10% of the
pages will be activated:
for i in `seq 0 100 10000000`; do echo $i 110; done > pattern-hot-10
iotrace.rb --load pattern-hot-10 --play /b/sparse
vmmon nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
and monitor /proc/vmstat during the time. The test box has 2G memory.
I carried out tests on fresh booted console as well as X desktop, and
fetched the vmstat numbers on
(1) begin: shortly after the big read IO starts;
(2) end: just before the big read IO stops;
(3) restore: the big read IO stops and the zsh working set restored
(4) restore X: after IO, switch back and forth between the urxvt and firefox
windows to restore their working set.
2.3) console mode results
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.29 VM_EXEC protection ON:
begin: 2481 2237 8694 630 0 574299
end: 275 231976 233914 633 776271 20933042
restore: 370 232154 234524 691 777183 20958453
2.6.29 VM_EXEC protection ON (second run):
begin: 2434 2237 8493 629 0 574195
end: 284 231970 233536 632 771918 20896129
restore: 399 232218 234789 690 774526 20957909
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 2479 2344 9659 210 0 579643
end: 284 232010 234142 260 772776 20917184
restore: 379 232159 234371 301 774888 20967849
The above console numbers show that
- The startup pgmajfault of 2.6.30-rc4-mm is merely 1/3 that of 2.6.29.
I'd attribute that improvement to the mmap readahead improvements :-)
- The pgmajfault increment during the file copy is 633-630=3 vs 260-210=50.
That's a huge improvement - which means with the VM_EXEC protection logic,
active mmap pages is pretty safe even under partially cache hot streaming IO.
- when active:inactive file lru size reaches 1:1, their scan rates is 1:20.8
under 10% cache hot IO. (computed with formula Dpgdeactivate:Dpgfree)
That roughly means the active mmap pages get 20.8 more chances to get
re-referenced to stay in memory.
- The absolute nr_mapped drops considerably to 1/9 during the big IO, and the
dropped pages are mostly inactive ones. The patch has almost no impact in
this aspect, that means it won't unnecessarily increase memory pressure.
(In contrast, your 20% mmap protection ratio will keep them all, and
therefore eliminate the extra 41 major faults to restore working set
of zsh etc.)
The iotrace.rb read throughput is
151.194384MB/s 284.198252s 100001x 450560b --load pattern-hot-10 --play /b/sparse
which means the inactive list is rotated at the speed of 250MB/s,
so a full scan of which takes about 3.5 seconds, while a full scan
of active file list takes about 77 seconds.
2.4) X mode results
We can reach roughly the same conclusions for X desktop:
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.30-rc4-mm VM_EXEC protection ON:
begin: 9740 8920 64075 561 0 678360
end: 768 218254 220029 565 798953 21057006
restore: 857 218543 220987 606 799462 21075710
restore X: 2414 218560 225344 797 799462 21080795
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 9368 5035 26389 554 0 633391
end: 770 218449 221230 661 646472 17832500
restore: 1113 218466 220978 710 649881 17905235
restore X: 2687 218650 225484 947 802700 21083584
- the absolute nr_mapped drops considerably (to 1/13 of the original size)
during the streaming IO.
- the delta of pgmajfault is 3 vs 107 during IO, or 236 vs 393
during the whole process.
Cc: Elladan <elladan@eskimo.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:12 +00:00
|
|
|
LIST_HEAD(l_active);
|
2008-10-19 03:26:14 +00:00
|
|
|
LIST_HEAD(l_inactive);
|
2017-02-22 23:44:18 +00:00
|
|
|
unsigned nr_deactivate, nr_activate;
|
|
|
|
unsigned nr_rotated = 0;
|
2012-05-29 22:06:53 +00:00
|
|
|
int file = is_file_lru(lru);
|
2016-07-28 22:45:31 +00:00
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(lruvec);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
lru_add_drain();
|
2011-11-01 00:06:55 +00:00
|
|
|
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
2012-01-13 01:18:15 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
|
2019-03-05 23:46:55 +00:00
|
|
|
&nr_scanned, sc, lru);
|
2012-01-13 01:17:50 +00:00
|
|
|
|
2016-07-28 22:45:31 +00:00
|
|
|
__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
|
2008-02-07 08:14:37 +00:00
|
|
|
|
2020-08-07 06:26:32 +00:00
|
|
|
if (!cgroup_reclaim(sc))
|
|
|
|
__count_vm_events(PGREFILL, nr_scanned);
|
2019-05-14 00:23:02 +00:00
|
|
|
__count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
|
mm: update_lru_size do the __mod_zone_page_state
Konstantin Khlebnikov pointed out (nearly four years ago, when lumpy
reclaim was removed) that lru_size can be updated by -nr_taken once per
call to isolate_lru_pages(), instead of page by page.
Update it inside isolate_lru_pages(), or at its two callsites? I chose
to update it at the callsites, rearranging and grouping the updates by
nr_taken and nr_scanned together in both.
With one exception, mem_cgroup_update_lru_size(,lru,) is then used where
__mod_zone_page_state(,NR_LRU_BASE+lru,) is used; and we shall be adding
some more calls in a future commit. Make the code a little smaller and
simpler by incorporating stat update in lru_size update.
The exception was move_active_pages_to_lru(), which aggregated the
pgmoved stat update separately from the individual lru_size updates; but
I still think this a simplification worth making.
However, the __mod_zone_page_state is not peculiar to mem_cgroups: so
better use the name update_lru_size, calls mem_cgroup_update_lru_size
when CONFIG_MEMCG.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Yang Shi <yang.shi@linaro.org>
Cc: Ning Qu <quning@gmail.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-05-20 00:12:38 +00:00
|
|
|
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
while (!list_empty(&l_hold)) {
|
2022-01-21 16:27:31 +00:00
|
|
|
struct folio *folio;
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
cond_resched();
|
2022-01-21 16:27:31 +00:00
|
|
|
folio = lru_to_folio(&l_hold);
|
|
|
|
list_del(&folio->lru);
|
2008-10-19 03:26:35 +00:00
|
|
|
|
2022-06-17 15:42:47 +00:00
|
|
|
if (unlikely(!folio_evictable(folio))) {
|
|
|
|
folio_putback_lru(folio);
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2012-03-21 23:34:00 +00:00
|
|
|
if (unlikely(buffer_heads_over_limit)) {
|
2022-09-02 19:26:39 +00:00
|
|
|
if (folio_test_private(folio) && folio_trylock(folio)) {
|
|
|
|
if (folio_test_private(folio))
|
2022-06-17 15:42:47 +00:00
|
|
|
filemap_release_folio(folio, 0);
|
|
|
|
folio_unlock(folio);
|
2012-03-21 23:34:00 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
mm: don't be stuck to rmap lock on reclaim path
The rmap locks(i_mmap_rwsem and anon_vma->root->rwsem) could be contended
under memory pressure if processes keep working on their vmas(e.g., fork,
mmap, munmap). It makes reclaim path stuck. In our real workload traces,
we see kswapd is waiting the lock for 300ms+(worst case, a sec) and it
makes other processes entering direct reclaim, which were also stuck on
the lock.
This patch makes lru aging path try_lock mode like shink_page_list so the
reclaim context will keep working with next lru pages without being stuck.
if it found the rmap lock contended, it rotates the page back to head of
lru in both active/inactive lrus to make them consistent behavior, which
is basic starting point rather than adding more heristic.
Since this patch introduces a new "contended" field as out-param along
with try_lock in-param in rmap_walk_control, it's not immutable any longer
if the try_lock is set so remove const keywords on rmap related functions.
Since rmap walking is already expensive operation, I doubt the const
would help sizable benefit( And we didn't have it until 5.17).
In a heavy app workload in Android, trace shows following statistics. It
almost removes rmap lock contention from reclaim path.
Martin Liu reported:
Before:
max_dur(ms) min_dur(ms) max-min(dur)ms avg_dur(ms) sum_dur(ms) count blocked_function
1632 0 1631 151.542173 31672 209 page_lock_anon_vma_read
601 0 601 145.544681 28817 198 rmap_walk_file
After:
max_dur(ms) min_dur(ms) max-min(dur)ms avg_dur(ms) sum_dur(ms) count blocked_function
NaN NaN NaN NaN NaN 0.0 NaN
0 0 0 0.127645 1 12 rmap_walk_file
[minchan@kernel.org: add comment, per Matthew]
Link: https://lkml.kernel.org/r/YnNqeB5tUf6LZ57b@google.com
Link: https://lkml.kernel.org/r/20220510215423.164547-1-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: John Dias <joaodias@google.com>
Cc: Tim Murray <timmurray@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Martin Liu <liumartin@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-19 21:08:54 +00:00
|
|
|
/* Referenced or rmap lock contention: rotate */
|
2022-01-21 16:27:31 +00:00
|
|
|
if (folio_referenced(folio, 0, sc->target_mem_cgroup,
|
mm: don't be stuck to rmap lock on reclaim path
The rmap locks(i_mmap_rwsem and anon_vma->root->rwsem) could be contended
under memory pressure if processes keep working on their vmas(e.g., fork,
mmap, munmap). It makes reclaim path stuck. In our real workload traces,
we see kswapd is waiting the lock for 300ms+(worst case, a sec) and it
makes other processes entering direct reclaim, which were also stuck on
the lock.
This patch makes lru aging path try_lock mode like shink_page_list so the
reclaim context will keep working with next lru pages without being stuck.
if it found the rmap lock contended, it rotates the page back to head of
lru in both active/inactive lrus to make them consistent behavior, which
is basic starting point rather than adding more heristic.
Since this patch introduces a new "contended" field as out-param along
with try_lock in-param in rmap_walk_control, it's not immutable any longer
if the try_lock is set so remove const keywords on rmap related functions.
Since rmap walking is already expensive operation, I doubt the const
would help sizable benefit( And we didn't have it until 5.17).
In a heavy app workload in Android, trace shows following statistics. It
almost removes rmap lock contention from reclaim path.
Martin Liu reported:
Before:
max_dur(ms) min_dur(ms) max-min(dur)ms avg_dur(ms) sum_dur(ms) count blocked_function
1632 0 1631 151.542173 31672 209 page_lock_anon_vma_read
601 0 601 145.544681 28817 198 rmap_walk_file
After:
max_dur(ms) min_dur(ms) max-min(dur)ms avg_dur(ms) sum_dur(ms) count blocked_function
NaN NaN NaN NaN NaN 0.0 NaN
0 0 0 0.127645 1 12 rmap_walk_file
[minchan@kernel.org: add comment, per Matthew]
Link: https://lkml.kernel.org/r/YnNqeB5tUf6LZ57b@google.com
Link: https://lkml.kernel.org/r/20220510215423.164547-1-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: John Dias <joaodias@google.com>
Cc: Tim Murray <timmurray@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Martin Liu <liumartin@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-19 21:08:54 +00:00
|
|
|
&vm_flags) != 0) {
|
vmscan: make mapped executable pages the first class citizen
Protect referenced PROT_EXEC mapped pages from being deactivated.
PROT_EXEC(or its internal presentation VM_EXEC) pages normally belong to some
currently running executables and their linked libraries, they shall really be
cached aggressively to provide good user experiences.
Thanks to Johannes Weiner for the advice to reuse the VMA walk in
page_referenced() to get the PROT_EXEC bit.
[more details]
( The consequences of this patch will have to be discussed together with
Rik van Riel's recent patch "vmscan: evict use-once pages first". )
( Some of the good points and insights are taken into this changelog.
Thanks to all the involved people for the great LKML discussions. )
the problem
===========
For a typical desktop, the most precious working set is composed of
*actively accessed*
(1) memory mapped executables
(2) and their anonymous pages
(3) and other files
(4) and the dcache/icache/.. slabs
while the least important data are
(5) infrequently used or use-once files
For a typical desktop, one major problem is busty and large amount of (5)
use-once files flushing out the working set.
Inside the working set, (4) dcache/icache have already been too sticky ;-)
So we only have to care (2) anonymous and (1)(3) file pages.
anonymous pages
===============
Anonymous pages are effectively immune to the streaming IO attack, because we
now have separate file/anon LRU lists. When the use-once files crowd into the
file LRU, the list's "quality" is significantly lowered. Therefore the scan
balance policy in get_scan_ratio() will choose to scan the (low quality) file
LRU much more frequently than the anon LRU.
file pages
==========
Rik proposed to *not* scan the active file LRU when the inactive list grows
larger than active list. This guarantees that when there are use-once streaming
IO, and the working set is not too large(so that active_size < inactive_size),
the active file LRU will *not* be scanned at all. So the not-too-large working
set can be well protected.
But there are also situations where the file working set is a bit large so that
(active_size >= inactive_size), or the streaming IOs are not purely use-once.
In these cases, the active list will be scanned slowly. Because the current
shrink_active_list() policy is to deactivate active pages regardless of their
referenced bits. The deactivated pages become susceptible to the streaming IO
attack: the inactive list could be scanned fast (500MB / 50MBps = 10s) so that
the deactivated pages don't have enough time to get re-referenced. Because a
user tend to switch between windows in intervals from seconds to minutes.
This patch holds mapped executable pages in the active list as long as they
are referenced during each full scan of the active list. Because the active
list is normally scanned much slower, they get longer grace time (eg. 100s)
for further references, which better matches the pace of user operations.
Therefore this patch greatly prolongs the in-cache time of executable code,
when there are moderate memory pressures.
before patch: guaranteed to be cached if reference intervals < I
after patch: guaranteed to be cached if reference intervals < I+A
(except when randomly reclaimed by the lumpy reclaim)
where
A = time to fully scan the active file LRU
I = time to fully scan the inactive file LRU
Note that normally A >> I.
side effects
============
This patch is safe in general, it restores the pre-2.6.28 mmap() behavior
but in a much smaller and well targeted scope.
One may worry about some one to abuse the PROT_EXEC heuristic. But as
Andrew Morton stated, there are other tricks to getting that sort of boost.
Another concern is the PROT_EXEC mapped pages growing large in rare cases,
and therefore hurting reclaim efficiency. But a sane application targeted for
large audience will never use PROT_EXEC for data mappings. If some home made
application tries to abuse that bit, it shall be aware of the consequences.
If it is abused to scale of 2/3 total memory, it gains nothing but overheads.
benchmarks
==========
1) memory tight desktop
1.1) brief summary
- clock time and major faults are reduced by 50%;
- pswpin numbers are reduced to ~1/3.
That means X desktop responsiveness is doubled under high memory/swap pressure.
1.2) test scenario
- nfsroot gnome desktop with 512M physical memory
- run some programs, and switch between the existing windows
after starting each new program.
1.3) progress timing (seconds)
before after programs
0.02 0.02 N xeyes
0.75 0.76 N firefox
2.02 1.88 N nautilus
3.36 3.17 N nautilus --browser
5.26 4.89 N gthumb
7.12 6.47 N gedit
9.22 8.16 N xpdf /usr/share/doc/shared-mime-info/shared-mime-info-spec.pdf
13.58 12.55 N xterm
15.87 14.57 N mlterm
18.63 17.06 N gnome-terminal
21.16 18.90 N urxvt
26.24 23.48 N gnome-system-monitor
28.72 26.52 N gnome-help
32.15 29.65 N gnome-dictionary
39.66 36.12 N /usr/games/sol
43.16 39.27 N /usr/games/gnometris
48.65 42.56 N /usr/games/gnect
53.31 47.03 N /usr/games/gtali
58.60 52.05 N /usr/games/iagno
65.77 55.42 N /usr/games/gnotravex
70.76 61.47 N /usr/games/mahjongg
76.15 67.11 N /usr/games/gnome-sudoku
86.32 75.15 N /usr/games/glines
92.21 79.70 N /usr/games/glchess
103.79 88.48 N /usr/games/gnomine
113.84 96.51 N /usr/games/gnotski
124.40 102.19 N /usr/games/gnibbles
137.41 114.93 N /usr/games/gnobots2
155.53 125.02 N /usr/games/blackjack
179.85 135.11 N /usr/games/same-gnome
224.49 154.50 N /usr/bin/gnome-window-properties
248.44 162.09 N /usr/bin/gnome-default-applications-properties
282.62 173.29 N /usr/bin/gnome-at-properties
323.72 188.21 N /usr/bin/gnome-typing-monitor
363.99 199.93 N /usr/bin/gnome-at-visual
394.21 206.95 N /usr/bin/gnome-sound-properties
435.14 224.49 N /usr/bin/gnome-at-mobility
463.05 234.11 N /usr/bin/gnome-keybinding-properties
503.75 248.59 N /usr/bin/gnome-about-me
554.00 276.27 N /usr/bin/gnome-display-properties
615.48 304.39 N /usr/bin/gnome-network-preferences
693.03 342.01 N /usr/bin/gnome-mouse-properties
759.90 388.58 N /usr/bin/gnome-appearance-properties
937.90 508.47 N /usr/bin/gnome-control-center
1109.75 587.57 N /usr/bin/gnome-keyboard-properties
1399.05 758.16 N : oocalc
1524.64 830.03 N : oodraw
1684.31 900.03 N : ooimpress
1874.04 993.91 N : oomath
2115.12 1081.89 N : ooweb
2369.02 1161.99 N : oowriter
Note that the last ": oo*" commands are actually commented out.
1.4) vmstat numbers (some relevant ones are marked with *)
before after
nr_free_pages 1293 3898
nr_inactive_anon 59956 53460
nr_active_anon 26815 30026
nr_inactive_file 2657 3218
nr_active_file 2019 2806
nr_unevictable 4 4
nr_mlock 4 4
nr_anon_pages 26706 27859
*nr_mapped 3542 4469
nr_file_pages 72232 67681
nr_dirty 1 0
nr_writeback 123 19
nr_slab_reclaimable 3375 3534
nr_slab_unreclaimable 11405 10665
nr_page_table_pages 8106 7864
nr_unstable 0 0
nr_bounce 0 0
*nr_vmscan_write 394776 230839
nr_writeback_temp 0 0
numa_hit 6843353 3318676
numa_miss 0 0
numa_foreign 0 0
numa_interleave 1719 1719
numa_local 6843353 3318676
numa_other 0 0
*pgpgin 5954683 2057175
*pgpgout 1578276 922744
*pswpin 1486615 512238
*pswpout 394568 230685
pgalloc_dma 277432 56602
pgalloc_dma32 6769477 3310348
pgalloc_normal 0 0
pgalloc_movable 0 0
pgfree 7048396 3371118
pgactivate 2036343 1471492
pgdeactivate 2189691 1612829
pgfault 3702176 3100702
*pgmajfault 452116 201343
pgrefill_dma 12185 7127
pgrefill_dma32 334384 653703
pgrefill_normal 0 0
pgrefill_movable 0 0
pgsteal_dma 74214 22179
pgsteal_dma32 3334164 1638029
pgsteal_normal 0 0
pgsteal_movable 0 0
pgscan_kswapd_dma 1081421 1216199
pgscan_kswapd_dma32 58979118 46002810
pgscan_kswapd_normal 0 0
pgscan_kswapd_movable 0 0
pgscan_direct_dma 2015438 1086109
pgscan_direct_dma32 55787823 36101597
pgscan_direct_normal 0 0
pgscan_direct_movable 0 0
pginodesteal 3461 7281
slabs_scanned 564864 527616
kswapd_steal 2889797 1448082
kswapd_inodesteal 14827 14835
pageoutrun 43459 21562
allocstall 9653 4032
pgrotated 384216 228631
1.5) free numbers at the end of the tests
before patch:
total used free shared buffers cached
Mem: 474 467 7 0 0 236
-/+ buffers/cache: 230 243
Swap: 1023 418 605
after patch:
total used free shared buffers cached
Mem: 474 457 16 0 0 236
-/+ buffers/cache: 221 253
Swap: 1023 404 619
2) memory flushing in a file server
2.1) brief summary
The number of major faults from 50 to 3 during 10% cache hot reads.
That means this patch successfully stops major faults when the active file
list is slowly scanned when there are partially cache hot streaming IO.
2.2) test scenario
Do 100000 pread(size=110 pages, offset=(i*100) pages), where 10% of the
pages will be activated:
for i in `seq 0 100 10000000`; do echo $i 110; done > pattern-hot-10
iotrace.rb --load pattern-hot-10 --play /b/sparse
vmmon nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
and monitor /proc/vmstat during the time. The test box has 2G memory.
I carried out tests on fresh booted console as well as X desktop, and
fetched the vmstat numbers on
(1) begin: shortly after the big read IO starts;
(2) end: just before the big read IO stops;
(3) restore: the big read IO stops and the zsh working set restored
(4) restore X: after IO, switch back and forth between the urxvt and firefox
windows to restore their working set.
2.3) console mode results
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.29 VM_EXEC protection ON:
begin: 2481 2237 8694 630 0 574299
end: 275 231976 233914 633 776271 20933042
restore: 370 232154 234524 691 777183 20958453
2.6.29 VM_EXEC protection ON (second run):
begin: 2434 2237 8493 629 0 574195
end: 284 231970 233536 632 771918 20896129
restore: 399 232218 234789 690 774526 20957909
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 2479 2344 9659 210 0 579643
end: 284 232010 234142 260 772776 20917184
restore: 379 232159 234371 301 774888 20967849
The above console numbers show that
- The startup pgmajfault of 2.6.30-rc4-mm is merely 1/3 that of 2.6.29.
I'd attribute that improvement to the mmap readahead improvements :-)
- The pgmajfault increment during the file copy is 633-630=3 vs 260-210=50.
That's a huge improvement - which means with the VM_EXEC protection logic,
active mmap pages is pretty safe even under partially cache hot streaming IO.
- when active:inactive file lru size reaches 1:1, their scan rates is 1:20.8
under 10% cache hot IO. (computed with formula Dpgdeactivate:Dpgfree)
That roughly means the active mmap pages get 20.8 more chances to get
re-referenced to stay in memory.
- The absolute nr_mapped drops considerably to 1/9 during the big IO, and the
dropped pages are mostly inactive ones. The patch has almost no impact in
this aspect, that means it won't unnecessarily increase memory pressure.
(In contrast, your 20% mmap protection ratio will keep them all, and
therefore eliminate the extra 41 major faults to restore working set
of zsh etc.)
The iotrace.rb read throughput is
151.194384MB/s 284.198252s 100001x 450560b --load pattern-hot-10 --play /b/sparse
which means the inactive list is rotated at the speed of 250MB/s,
so a full scan of which takes about 3.5 seconds, while a full scan
of active file list takes about 77 seconds.
2.4) X mode results
We can reach roughly the same conclusions for X desktop:
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.30-rc4-mm VM_EXEC protection ON:
begin: 9740 8920 64075 561 0 678360
end: 768 218254 220029 565 798953 21057006
restore: 857 218543 220987 606 799462 21075710
restore X: 2414 218560 225344 797 799462 21080795
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 9368 5035 26389 554 0 633391
end: 770 218449 221230 661 646472 17832500
restore: 1113 218466 220978 710 649881 17905235
restore X: 2687 218650 225484 947 802700 21083584
- the absolute nr_mapped drops considerably (to 1/13 of the original size)
during the streaming IO.
- the delta of pgmajfault is 3 vs 107 during IO, or 236 vs 393
during the whole process.
Cc: Elladan <elladan@eskimo.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:12 +00:00
|
|
|
/*
|
2022-06-17 15:42:47 +00:00
|
|
|
* Identify referenced, file-backed active folios and
|
vmscan: make mapped executable pages the first class citizen
Protect referenced PROT_EXEC mapped pages from being deactivated.
PROT_EXEC(or its internal presentation VM_EXEC) pages normally belong to some
currently running executables and their linked libraries, they shall really be
cached aggressively to provide good user experiences.
Thanks to Johannes Weiner for the advice to reuse the VMA walk in
page_referenced() to get the PROT_EXEC bit.
[more details]
( The consequences of this patch will have to be discussed together with
Rik van Riel's recent patch "vmscan: evict use-once pages first". )
( Some of the good points and insights are taken into this changelog.
Thanks to all the involved people for the great LKML discussions. )
the problem
===========
For a typical desktop, the most precious working set is composed of
*actively accessed*
(1) memory mapped executables
(2) and their anonymous pages
(3) and other files
(4) and the dcache/icache/.. slabs
while the least important data are
(5) infrequently used or use-once files
For a typical desktop, one major problem is busty and large amount of (5)
use-once files flushing out the working set.
Inside the working set, (4) dcache/icache have already been too sticky ;-)
So we only have to care (2) anonymous and (1)(3) file pages.
anonymous pages
===============
Anonymous pages are effectively immune to the streaming IO attack, because we
now have separate file/anon LRU lists. When the use-once files crowd into the
file LRU, the list's "quality" is significantly lowered. Therefore the scan
balance policy in get_scan_ratio() will choose to scan the (low quality) file
LRU much more frequently than the anon LRU.
file pages
==========
Rik proposed to *not* scan the active file LRU when the inactive list grows
larger than active list. This guarantees that when there are use-once streaming
IO, and the working set is not too large(so that active_size < inactive_size),
the active file LRU will *not* be scanned at all. So the not-too-large working
set can be well protected.
But there are also situations where the file working set is a bit large so that
(active_size >= inactive_size), or the streaming IOs are not purely use-once.
In these cases, the active list will be scanned slowly. Because the current
shrink_active_list() policy is to deactivate active pages regardless of their
referenced bits. The deactivated pages become susceptible to the streaming IO
attack: the inactive list could be scanned fast (500MB / 50MBps = 10s) so that
the deactivated pages don't have enough time to get re-referenced. Because a
user tend to switch between windows in intervals from seconds to minutes.
This patch holds mapped executable pages in the active list as long as they
are referenced during each full scan of the active list. Because the active
list is normally scanned much slower, they get longer grace time (eg. 100s)
for further references, which better matches the pace of user operations.
Therefore this patch greatly prolongs the in-cache time of executable code,
when there are moderate memory pressures.
before patch: guaranteed to be cached if reference intervals < I
after patch: guaranteed to be cached if reference intervals < I+A
(except when randomly reclaimed by the lumpy reclaim)
where
A = time to fully scan the active file LRU
I = time to fully scan the inactive file LRU
Note that normally A >> I.
side effects
============
This patch is safe in general, it restores the pre-2.6.28 mmap() behavior
but in a much smaller and well targeted scope.
One may worry about some one to abuse the PROT_EXEC heuristic. But as
Andrew Morton stated, there are other tricks to getting that sort of boost.
Another concern is the PROT_EXEC mapped pages growing large in rare cases,
and therefore hurting reclaim efficiency. But a sane application targeted for
large audience will never use PROT_EXEC for data mappings. If some home made
application tries to abuse that bit, it shall be aware of the consequences.
If it is abused to scale of 2/3 total memory, it gains nothing but overheads.
benchmarks
==========
1) memory tight desktop
1.1) brief summary
- clock time and major faults are reduced by 50%;
- pswpin numbers are reduced to ~1/3.
That means X desktop responsiveness is doubled under high memory/swap pressure.
1.2) test scenario
- nfsroot gnome desktop with 512M physical memory
- run some programs, and switch between the existing windows
after starting each new program.
1.3) progress timing (seconds)
before after programs
0.02 0.02 N xeyes
0.75 0.76 N firefox
2.02 1.88 N nautilus
3.36 3.17 N nautilus --browser
5.26 4.89 N gthumb
7.12 6.47 N gedit
9.22 8.16 N xpdf /usr/share/doc/shared-mime-info/shared-mime-info-spec.pdf
13.58 12.55 N xterm
15.87 14.57 N mlterm
18.63 17.06 N gnome-terminal
21.16 18.90 N urxvt
26.24 23.48 N gnome-system-monitor
28.72 26.52 N gnome-help
32.15 29.65 N gnome-dictionary
39.66 36.12 N /usr/games/sol
43.16 39.27 N /usr/games/gnometris
48.65 42.56 N /usr/games/gnect
53.31 47.03 N /usr/games/gtali
58.60 52.05 N /usr/games/iagno
65.77 55.42 N /usr/games/gnotravex
70.76 61.47 N /usr/games/mahjongg
76.15 67.11 N /usr/games/gnome-sudoku
86.32 75.15 N /usr/games/glines
92.21 79.70 N /usr/games/glchess
103.79 88.48 N /usr/games/gnomine
113.84 96.51 N /usr/games/gnotski
124.40 102.19 N /usr/games/gnibbles
137.41 114.93 N /usr/games/gnobots2
155.53 125.02 N /usr/games/blackjack
179.85 135.11 N /usr/games/same-gnome
224.49 154.50 N /usr/bin/gnome-window-properties
248.44 162.09 N /usr/bin/gnome-default-applications-properties
282.62 173.29 N /usr/bin/gnome-at-properties
323.72 188.21 N /usr/bin/gnome-typing-monitor
363.99 199.93 N /usr/bin/gnome-at-visual
394.21 206.95 N /usr/bin/gnome-sound-properties
435.14 224.49 N /usr/bin/gnome-at-mobility
463.05 234.11 N /usr/bin/gnome-keybinding-properties
503.75 248.59 N /usr/bin/gnome-about-me
554.00 276.27 N /usr/bin/gnome-display-properties
615.48 304.39 N /usr/bin/gnome-network-preferences
693.03 342.01 N /usr/bin/gnome-mouse-properties
759.90 388.58 N /usr/bin/gnome-appearance-properties
937.90 508.47 N /usr/bin/gnome-control-center
1109.75 587.57 N /usr/bin/gnome-keyboard-properties
1399.05 758.16 N : oocalc
1524.64 830.03 N : oodraw
1684.31 900.03 N : ooimpress
1874.04 993.91 N : oomath
2115.12 1081.89 N : ooweb
2369.02 1161.99 N : oowriter
Note that the last ": oo*" commands are actually commented out.
1.4) vmstat numbers (some relevant ones are marked with *)
before after
nr_free_pages 1293 3898
nr_inactive_anon 59956 53460
nr_active_anon 26815 30026
nr_inactive_file 2657 3218
nr_active_file 2019 2806
nr_unevictable 4 4
nr_mlock 4 4
nr_anon_pages 26706 27859
*nr_mapped 3542 4469
nr_file_pages 72232 67681
nr_dirty 1 0
nr_writeback 123 19
nr_slab_reclaimable 3375 3534
nr_slab_unreclaimable 11405 10665
nr_page_table_pages 8106 7864
nr_unstable 0 0
nr_bounce 0 0
*nr_vmscan_write 394776 230839
nr_writeback_temp 0 0
numa_hit 6843353 3318676
numa_miss 0 0
numa_foreign 0 0
numa_interleave 1719 1719
numa_local 6843353 3318676
numa_other 0 0
*pgpgin 5954683 2057175
*pgpgout 1578276 922744
*pswpin 1486615 512238
*pswpout 394568 230685
pgalloc_dma 277432 56602
pgalloc_dma32 6769477 3310348
pgalloc_normal 0 0
pgalloc_movable 0 0
pgfree 7048396 3371118
pgactivate 2036343 1471492
pgdeactivate 2189691 1612829
pgfault 3702176 3100702
*pgmajfault 452116 201343
pgrefill_dma 12185 7127
pgrefill_dma32 334384 653703
pgrefill_normal 0 0
pgrefill_movable 0 0
pgsteal_dma 74214 22179
pgsteal_dma32 3334164 1638029
pgsteal_normal 0 0
pgsteal_movable 0 0
pgscan_kswapd_dma 1081421 1216199
pgscan_kswapd_dma32 58979118 46002810
pgscan_kswapd_normal 0 0
pgscan_kswapd_movable 0 0
pgscan_direct_dma 2015438 1086109
pgscan_direct_dma32 55787823 36101597
pgscan_direct_normal 0 0
pgscan_direct_movable 0 0
pginodesteal 3461 7281
slabs_scanned 564864 527616
kswapd_steal 2889797 1448082
kswapd_inodesteal 14827 14835
pageoutrun 43459 21562
allocstall 9653 4032
pgrotated 384216 228631
1.5) free numbers at the end of the tests
before patch:
total used free shared buffers cached
Mem: 474 467 7 0 0 236
-/+ buffers/cache: 230 243
Swap: 1023 418 605
after patch:
total used free shared buffers cached
Mem: 474 457 16 0 0 236
-/+ buffers/cache: 221 253
Swap: 1023 404 619
2) memory flushing in a file server
2.1) brief summary
The number of major faults from 50 to 3 during 10% cache hot reads.
That means this patch successfully stops major faults when the active file
list is slowly scanned when there are partially cache hot streaming IO.
2.2) test scenario
Do 100000 pread(size=110 pages, offset=(i*100) pages), where 10% of the
pages will be activated:
for i in `seq 0 100 10000000`; do echo $i 110; done > pattern-hot-10
iotrace.rb --load pattern-hot-10 --play /b/sparse
vmmon nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
and monitor /proc/vmstat during the time. The test box has 2G memory.
I carried out tests on fresh booted console as well as X desktop, and
fetched the vmstat numbers on
(1) begin: shortly after the big read IO starts;
(2) end: just before the big read IO stops;
(3) restore: the big read IO stops and the zsh working set restored
(4) restore X: after IO, switch back and forth between the urxvt and firefox
windows to restore their working set.
2.3) console mode results
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.29 VM_EXEC protection ON:
begin: 2481 2237 8694 630 0 574299
end: 275 231976 233914 633 776271 20933042
restore: 370 232154 234524 691 777183 20958453
2.6.29 VM_EXEC protection ON (second run):
begin: 2434 2237 8493 629 0 574195
end: 284 231970 233536 632 771918 20896129
restore: 399 232218 234789 690 774526 20957909
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 2479 2344 9659 210 0 579643
end: 284 232010 234142 260 772776 20917184
restore: 379 232159 234371 301 774888 20967849
The above console numbers show that
- The startup pgmajfault of 2.6.30-rc4-mm is merely 1/3 that of 2.6.29.
I'd attribute that improvement to the mmap readahead improvements :-)
- The pgmajfault increment during the file copy is 633-630=3 vs 260-210=50.
That's a huge improvement - which means with the VM_EXEC protection logic,
active mmap pages is pretty safe even under partially cache hot streaming IO.
- when active:inactive file lru size reaches 1:1, their scan rates is 1:20.8
under 10% cache hot IO. (computed with formula Dpgdeactivate:Dpgfree)
That roughly means the active mmap pages get 20.8 more chances to get
re-referenced to stay in memory.
- The absolute nr_mapped drops considerably to 1/9 during the big IO, and the
dropped pages are mostly inactive ones. The patch has almost no impact in
this aspect, that means it won't unnecessarily increase memory pressure.
(In contrast, your 20% mmap protection ratio will keep them all, and
therefore eliminate the extra 41 major faults to restore working set
of zsh etc.)
The iotrace.rb read throughput is
151.194384MB/s 284.198252s 100001x 450560b --load pattern-hot-10 --play /b/sparse
which means the inactive list is rotated at the speed of 250MB/s,
so a full scan of which takes about 3.5 seconds, while a full scan
of active file list takes about 77 seconds.
2.4) X mode results
We can reach roughly the same conclusions for X desktop:
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.30-rc4-mm VM_EXEC protection ON:
begin: 9740 8920 64075 561 0 678360
end: 768 218254 220029 565 798953 21057006
restore: 857 218543 220987 606 799462 21075710
restore X: 2414 218560 225344 797 799462 21080795
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 9368 5035 26389 554 0 633391
end: 770 218449 221230 661 646472 17832500
restore: 1113 218466 220978 710 649881 17905235
restore X: 2687 218650 225484 947 802700 21083584
- the absolute nr_mapped drops considerably (to 1/13 of the original size)
during the streaming IO.
- the delta of pgmajfault is 3 vs 107 during IO, or 236 vs 393
during the whole process.
Cc: Elladan <elladan@eskimo.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:12 +00:00
|
|
|
* give them one more trip around the active list. So
|
|
|
|
* that executable code get better chances to stay in
|
2022-06-17 15:42:47 +00:00
|
|
|
* memory under moderate memory pressure. Anon folios
|
vmscan: make mapped executable pages the first class citizen
Protect referenced PROT_EXEC mapped pages from being deactivated.
PROT_EXEC(or its internal presentation VM_EXEC) pages normally belong to some
currently running executables and their linked libraries, they shall really be
cached aggressively to provide good user experiences.
Thanks to Johannes Weiner for the advice to reuse the VMA walk in
page_referenced() to get the PROT_EXEC bit.
[more details]
( The consequences of this patch will have to be discussed together with
Rik van Riel's recent patch "vmscan: evict use-once pages first". )
( Some of the good points and insights are taken into this changelog.
Thanks to all the involved people for the great LKML discussions. )
the problem
===========
For a typical desktop, the most precious working set is composed of
*actively accessed*
(1) memory mapped executables
(2) and their anonymous pages
(3) and other files
(4) and the dcache/icache/.. slabs
while the least important data are
(5) infrequently used or use-once files
For a typical desktop, one major problem is busty and large amount of (5)
use-once files flushing out the working set.
Inside the working set, (4) dcache/icache have already been too sticky ;-)
So we only have to care (2) anonymous and (1)(3) file pages.
anonymous pages
===============
Anonymous pages are effectively immune to the streaming IO attack, because we
now have separate file/anon LRU lists. When the use-once files crowd into the
file LRU, the list's "quality" is significantly lowered. Therefore the scan
balance policy in get_scan_ratio() will choose to scan the (low quality) file
LRU much more frequently than the anon LRU.
file pages
==========
Rik proposed to *not* scan the active file LRU when the inactive list grows
larger than active list. This guarantees that when there are use-once streaming
IO, and the working set is not too large(so that active_size < inactive_size),
the active file LRU will *not* be scanned at all. So the not-too-large working
set can be well protected.
But there are also situations where the file working set is a bit large so that
(active_size >= inactive_size), or the streaming IOs are not purely use-once.
In these cases, the active list will be scanned slowly. Because the current
shrink_active_list() policy is to deactivate active pages regardless of their
referenced bits. The deactivated pages become susceptible to the streaming IO
attack: the inactive list could be scanned fast (500MB / 50MBps = 10s) so that
the deactivated pages don't have enough time to get re-referenced. Because a
user tend to switch between windows in intervals from seconds to minutes.
This patch holds mapped executable pages in the active list as long as they
are referenced during each full scan of the active list. Because the active
list is normally scanned much slower, they get longer grace time (eg. 100s)
for further references, which better matches the pace of user operations.
Therefore this patch greatly prolongs the in-cache time of executable code,
when there are moderate memory pressures.
before patch: guaranteed to be cached if reference intervals < I
after patch: guaranteed to be cached if reference intervals < I+A
(except when randomly reclaimed by the lumpy reclaim)
where
A = time to fully scan the active file LRU
I = time to fully scan the inactive file LRU
Note that normally A >> I.
side effects
============
This patch is safe in general, it restores the pre-2.6.28 mmap() behavior
but in a much smaller and well targeted scope.
One may worry about some one to abuse the PROT_EXEC heuristic. But as
Andrew Morton stated, there are other tricks to getting that sort of boost.
Another concern is the PROT_EXEC mapped pages growing large in rare cases,
and therefore hurting reclaim efficiency. But a sane application targeted for
large audience will never use PROT_EXEC for data mappings. If some home made
application tries to abuse that bit, it shall be aware of the consequences.
If it is abused to scale of 2/3 total memory, it gains nothing but overheads.
benchmarks
==========
1) memory tight desktop
1.1) brief summary
- clock time and major faults are reduced by 50%;
- pswpin numbers are reduced to ~1/3.
That means X desktop responsiveness is doubled under high memory/swap pressure.
1.2) test scenario
- nfsroot gnome desktop with 512M physical memory
- run some programs, and switch between the existing windows
after starting each new program.
1.3) progress timing (seconds)
before after programs
0.02 0.02 N xeyes
0.75 0.76 N firefox
2.02 1.88 N nautilus
3.36 3.17 N nautilus --browser
5.26 4.89 N gthumb
7.12 6.47 N gedit
9.22 8.16 N xpdf /usr/share/doc/shared-mime-info/shared-mime-info-spec.pdf
13.58 12.55 N xterm
15.87 14.57 N mlterm
18.63 17.06 N gnome-terminal
21.16 18.90 N urxvt
26.24 23.48 N gnome-system-monitor
28.72 26.52 N gnome-help
32.15 29.65 N gnome-dictionary
39.66 36.12 N /usr/games/sol
43.16 39.27 N /usr/games/gnometris
48.65 42.56 N /usr/games/gnect
53.31 47.03 N /usr/games/gtali
58.60 52.05 N /usr/games/iagno
65.77 55.42 N /usr/games/gnotravex
70.76 61.47 N /usr/games/mahjongg
76.15 67.11 N /usr/games/gnome-sudoku
86.32 75.15 N /usr/games/glines
92.21 79.70 N /usr/games/glchess
103.79 88.48 N /usr/games/gnomine
113.84 96.51 N /usr/games/gnotski
124.40 102.19 N /usr/games/gnibbles
137.41 114.93 N /usr/games/gnobots2
155.53 125.02 N /usr/games/blackjack
179.85 135.11 N /usr/games/same-gnome
224.49 154.50 N /usr/bin/gnome-window-properties
248.44 162.09 N /usr/bin/gnome-default-applications-properties
282.62 173.29 N /usr/bin/gnome-at-properties
323.72 188.21 N /usr/bin/gnome-typing-monitor
363.99 199.93 N /usr/bin/gnome-at-visual
394.21 206.95 N /usr/bin/gnome-sound-properties
435.14 224.49 N /usr/bin/gnome-at-mobility
463.05 234.11 N /usr/bin/gnome-keybinding-properties
503.75 248.59 N /usr/bin/gnome-about-me
554.00 276.27 N /usr/bin/gnome-display-properties
615.48 304.39 N /usr/bin/gnome-network-preferences
693.03 342.01 N /usr/bin/gnome-mouse-properties
759.90 388.58 N /usr/bin/gnome-appearance-properties
937.90 508.47 N /usr/bin/gnome-control-center
1109.75 587.57 N /usr/bin/gnome-keyboard-properties
1399.05 758.16 N : oocalc
1524.64 830.03 N : oodraw
1684.31 900.03 N : ooimpress
1874.04 993.91 N : oomath
2115.12 1081.89 N : ooweb
2369.02 1161.99 N : oowriter
Note that the last ": oo*" commands are actually commented out.
1.4) vmstat numbers (some relevant ones are marked with *)
before after
nr_free_pages 1293 3898
nr_inactive_anon 59956 53460
nr_active_anon 26815 30026
nr_inactive_file 2657 3218
nr_active_file 2019 2806
nr_unevictable 4 4
nr_mlock 4 4
nr_anon_pages 26706 27859
*nr_mapped 3542 4469
nr_file_pages 72232 67681
nr_dirty 1 0
nr_writeback 123 19
nr_slab_reclaimable 3375 3534
nr_slab_unreclaimable 11405 10665
nr_page_table_pages 8106 7864
nr_unstable 0 0
nr_bounce 0 0
*nr_vmscan_write 394776 230839
nr_writeback_temp 0 0
numa_hit 6843353 3318676
numa_miss 0 0
numa_foreign 0 0
numa_interleave 1719 1719
numa_local 6843353 3318676
numa_other 0 0
*pgpgin 5954683 2057175
*pgpgout 1578276 922744
*pswpin 1486615 512238
*pswpout 394568 230685
pgalloc_dma 277432 56602
pgalloc_dma32 6769477 3310348
pgalloc_normal 0 0
pgalloc_movable 0 0
pgfree 7048396 3371118
pgactivate 2036343 1471492
pgdeactivate 2189691 1612829
pgfault 3702176 3100702
*pgmajfault 452116 201343
pgrefill_dma 12185 7127
pgrefill_dma32 334384 653703
pgrefill_normal 0 0
pgrefill_movable 0 0
pgsteal_dma 74214 22179
pgsteal_dma32 3334164 1638029
pgsteal_normal 0 0
pgsteal_movable 0 0
pgscan_kswapd_dma 1081421 1216199
pgscan_kswapd_dma32 58979118 46002810
pgscan_kswapd_normal 0 0
pgscan_kswapd_movable 0 0
pgscan_direct_dma 2015438 1086109
pgscan_direct_dma32 55787823 36101597
pgscan_direct_normal 0 0
pgscan_direct_movable 0 0
pginodesteal 3461 7281
slabs_scanned 564864 527616
kswapd_steal 2889797 1448082
kswapd_inodesteal 14827 14835
pageoutrun 43459 21562
allocstall 9653 4032
pgrotated 384216 228631
1.5) free numbers at the end of the tests
before patch:
total used free shared buffers cached
Mem: 474 467 7 0 0 236
-/+ buffers/cache: 230 243
Swap: 1023 418 605
after patch:
total used free shared buffers cached
Mem: 474 457 16 0 0 236
-/+ buffers/cache: 221 253
Swap: 1023 404 619
2) memory flushing in a file server
2.1) brief summary
The number of major faults from 50 to 3 during 10% cache hot reads.
That means this patch successfully stops major faults when the active file
list is slowly scanned when there are partially cache hot streaming IO.
2.2) test scenario
Do 100000 pread(size=110 pages, offset=(i*100) pages), where 10% of the
pages will be activated:
for i in `seq 0 100 10000000`; do echo $i 110; done > pattern-hot-10
iotrace.rb --load pattern-hot-10 --play /b/sparse
vmmon nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
and monitor /proc/vmstat during the time. The test box has 2G memory.
I carried out tests on fresh booted console as well as X desktop, and
fetched the vmstat numbers on
(1) begin: shortly after the big read IO starts;
(2) end: just before the big read IO stops;
(3) restore: the big read IO stops and the zsh working set restored
(4) restore X: after IO, switch back and forth between the urxvt and firefox
windows to restore their working set.
2.3) console mode results
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.29 VM_EXEC protection ON:
begin: 2481 2237 8694 630 0 574299
end: 275 231976 233914 633 776271 20933042
restore: 370 232154 234524 691 777183 20958453
2.6.29 VM_EXEC protection ON (second run):
begin: 2434 2237 8493 629 0 574195
end: 284 231970 233536 632 771918 20896129
restore: 399 232218 234789 690 774526 20957909
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 2479 2344 9659 210 0 579643
end: 284 232010 234142 260 772776 20917184
restore: 379 232159 234371 301 774888 20967849
The above console numbers show that
- The startup pgmajfault of 2.6.30-rc4-mm is merely 1/3 that of 2.6.29.
I'd attribute that improvement to the mmap readahead improvements :-)
- The pgmajfault increment during the file copy is 633-630=3 vs 260-210=50.
That's a huge improvement - which means with the VM_EXEC protection logic,
active mmap pages is pretty safe even under partially cache hot streaming IO.
- when active:inactive file lru size reaches 1:1, their scan rates is 1:20.8
under 10% cache hot IO. (computed with formula Dpgdeactivate:Dpgfree)
That roughly means the active mmap pages get 20.8 more chances to get
re-referenced to stay in memory.
- The absolute nr_mapped drops considerably to 1/9 during the big IO, and the
dropped pages are mostly inactive ones. The patch has almost no impact in
this aspect, that means it won't unnecessarily increase memory pressure.
(In contrast, your 20% mmap protection ratio will keep them all, and
therefore eliminate the extra 41 major faults to restore working set
of zsh etc.)
The iotrace.rb read throughput is
151.194384MB/s 284.198252s 100001x 450560b --load pattern-hot-10 --play /b/sparse
which means the inactive list is rotated at the speed of 250MB/s,
so a full scan of which takes about 3.5 seconds, while a full scan
of active file list takes about 77 seconds.
2.4) X mode results
We can reach roughly the same conclusions for X desktop:
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.30-rc4-mm VM_EXEC protection ON:
begin: 9740 8920 64075 561 0 678360
end: 768 218254 220029 565 798953 21057006
restore: 857 218543 220987 606 799462 21075710
restore X: 2414 218560 225344 797 799462 21080795
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 9368 5035 26389 554 0 633391
end: 770 218449 221230 661 646472 17832500
restore: 1113 218466 220978 710 649881 17905235
restore X: 2687 218650 225484 947 802700 21083584
- the absolute nr_mapped drops considerably (to 1/13 of the original size)
during the streaming IO.
- the delta of pgmajfault is 3 vs 107 during IO, or 236 vs 393
during the whole process.
Cc: Elladan <elladan@eskimo.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:12 +00:00
|
|
|
* are not likely to be evicted by use-once streaming
|
2022-06-17 15:42:47 +00:00
|
|
|
* IO, plus JVM can create lots of anon VM_EXEC folios,
|
vmscan: make mapped executable pages the first class citizen
Protect referenced PROT_EXEC mapped pages from being deactivated.
PROT_EXEC(or its internal presentation VM_EXEC) pages normally belong to some
currently running executables and their linked libraries, they shall really be
cached aggressively to provide good user experiences.
Thanks to Johannes Weiner for the advice to reuse the VMA walk in
page_referenced() to get the PROT_EXEC bit.
[more details]
( The consequences of this patch will have to be discussed together with
Rik van Riel's recent patch "vmscan: evict use-once pages first". )
( Some of the good points and insights are taken into this changelog.
Thanks to all the involved people for the great LKML discussions. )
the problem
===========
For a typical desktop, the most precious working set is composed of
*actively accessed*
(1) memory mapped executables
(2) and their anonymous pages
(3) and other files
(4) and the dcache/icache/.. slabs
while the least important data are
(5) infrequently used or use-once files
For a typical desktop, one major problem is busty and large amount of (5)
use-once files flushing out the working set.
Inside the working set, (4) dcache/icache have already been too sticky ;-)
So we only have to care (2) anonymous and (1)(3) file pages.
anonymous pages
===============
Anonymous pages are effectively immune to the streaming IO attack, because we
now have separate file/anon LRU lists. When the use-once files crowd into the
file LRU, the list's "quality" is significantly lowered. Therefore the scan
balance policy in get_scan_ratio() will choose to scan the (low quality) file
LRU much more frequently than the anon LRU.
file pages
==========
Rik proposed to *not* scan the active file LRU when the inactive list grows
larger than active list. This guarantees that when there are use-once streaming
IO, and the working set is not too large(so that active_size < inactive_size),
the active file LRU will *not* be scanned at all. So the not-too-large working
set can be well protected.
But there are also situations where the file working set is a bit large so that
(active_size >= inactive_size), or the streaming IOs are not purely use-once.
In these cases, the active list will be scanned slowly. Because the current
shrink_active_list() policy is to deactivate active pages regardless of their
referenced bits. The deactivated pages become susceptible to the streaming IO
attack: the inactive list could be scanned fast (500MB / 50MBps = 10s) so that
the deactivated pages don't have enough time to get re-referenced. Because a
user tend to switch between windows in intervals from seconds to minutes.
This patch holds mapped executable pages in the active list as long as they
are referenced during each full scan of the active list. Because the active
list is normally scanned much slower, they get longer grace time (eg. 100s)
for further references, which better matches the pace of user operations.
Therefore this patch greatly prolongs the in-cache time of executable code,
when there are moderate memory pressures.
before patch: guaranteed to be cached if reference intervals < I
after patch: guaranteed to be cached if reference intervals < I+A
(except when randomly reclaimed by the lumpy reclaim)
where
A = time to fully scan the active file LRU
I = time to fully scan the inactive file LRU
Note that normally A >> I.
side effects
============
This patch is safe in general, it restores the pre-2.6.28 mmap() behavior
but in a much smaller and well targeted scope.
One may worry about some one to abuse the PROT_EXEC heuristic. But as
Andrew Morton stated, there are other tricks to getting that sort of boost.
Another concern is the PROT_EXEC mapped pages growing large in rare cases,
and therefore hurting reclaim efficiency. But a sane application targeted for
large audience will never use PROT_EXEC for data mappings. If some home made
application tries to abuse that bit, it shall be aware of the consequences.
If it is abused to scale of 2/3 total memory, it gains nothing but overheads.
benchmarks
==========
1) memory tight desktop
1.1) brief summary
- clock time and major faults are reduced by 50%;
- pswpin numbers are reduced to ~1/3.
That means X desktop responsiveness is doubled under high memory/swap pressure.
1.2) test scenario
- nfsroot gnome desktop with 512M physical memory
- run some programs, and switch between the existing windows
after starting each new program.
1.3) progress timing (seconds)
before after programs
0.02 0.02 N xeyes
0.75 0.76 N firefox
2.02 1.88 N nautilus
3.36 3.17 N nautilus --browser
5.26 4.89 N gthumb
7.12 6.47 N gedit
9.22 8.16 N xpdf /usr/share/doc/shared-mime-info/shared-mime-info-spec.pdf
13.58 12.55 N xterm
15.87 14.57 N mlterm
18.63 17.06 N gnome-terminal
21.16 18.90 N urxvt
26.24 23.48 N gnome-system-monitor
28.72 26.52 N gnome-help
32.15 29.65 N gnome-dictionary
39.66 36.12 N /usr/games/sol
43.16 39.27 N /usr/games/gnometris
48.65 42.56 N /usr/games/gnect
53.31 47.03 N /usr/games/gtali
58.60 52.05 N /usr/games/iagno
65.77 55.42 N /usr/games/gnotravex
70.76 61.47 N /usr/games/mahjongg
76.15 67.11 N /usr/games/gnome-sudoku
86.32 75.15 N /usr/games/glines
92.21 79.70 N /usr/games/glchess
103.79 88.48 N /usr/games/gnomine
113.84 96.51 N /usr/games/gnotski
124.40 102.19 N /usr/games/gnibbles
137.41 114.93 N /usr/games/gnobots2
155.53 125.02 N /usr/games/blackjack
179.85 135.11 N /usr/games/same-gnome
224.49 154.50 N /usr/bin/gnome-window-properties
248.44 162.09 N /usr/bin/gnome-default-applications-properties
282.62 173.29 N /usr/bin/gnome-at-properties
323.72 188.21 N /usr/bin/gnome-typing-monitor
363.99 199.93 N /usr/bin/gnome-at-visual
394.21 206.95 N /usr/bin/gnome-sound-properties
435.14 224.49 N /usr/bin/gnome-at-mobility
463.05 234.11 N /usr/bin/gnome-keybinding-properties
503.75 248.59 N /usr/bin/gnome-about-me
554.00 276.27 N /usr/bin/gnome-display-properties
615.48 304.39 N /usr/bin/gnome-network-preferences
693.03 342.01 N /usr/bin/gnome-mouse-properties
759.90 388.58 N /usr/bin/gnome-appearance-properties
937.90 508.47 N /usr/bin/gnome-control-center
1109.75 587.57 N /usr/bin/gnome-keyboard-properties
1399.05 758.16 N : oocalc
1524.64 830.03 N : oodraw
1684.31 900.03 N : ooimpress
1874.04 993.91 N : oomath
2115.12 1081.89 N : ooweb
2369.02 1161.99 N : oowriter
Note that the last ": oo*" commands are actually commented out.
1.4) vmstat numbers (some relevant ones are marked with *)
before after
nr_free_pages 1293 3898
nr_inactive_anon 59956 53460
nr_active_anon 26815 30026
nr_inactive_file 2657 3218
nr_active_file 2019 2806
nr_unevictable 4 4
nr_mlock 4 4
nr_anon_pages 26706 27859
*nr_mapped 3542 4469
nr_file_pages 72232 67681
nr_dirty 1 0
nr_writeback 123 19
nr_slab_reclaimable 3375 3534
nr_slab_unreclaimable 11405 10665
nr_page_table_pages 8106 7864
nr_unstable 0 0
nr_bounce 0 0
*nr_vmscan_write 394776 230839
nr_writeback_temp 0 0
numa_hit 6843353 3318676
numa_miss 0 0
numa_foreign 0 0
numa_interleave 1719 1719
numa_local 6843353 3318676
numa_other 0 0
*pgpgin 5954683 2057175
*pgpgout 1578276 922744
*pswpin 1486615 512238
*pswpout 394568 230685
pgalloc_dma 277432 56602
pgalloc_dma32 6769477 3310348
pgalloc_normal 0 0
pgalloc_movable 0 0
pgfree 7048396 3371118
pgactivate 2036343 1471492
pgdeactivate 2189691 1612829
pgfault 3702176 3100702
*pgmajfault 452116 201343
pgrefill_dma 12185 7127
pgrefill_dma32 334384 653703
pgrefill_normal 0 0
pgrefill_movable 0 0
pgsteal_dma 74214 22179
pgsteal_dma32 3334164 1638029
pgsteal_normal 0 0
pgsteal_movable 0 0
pgscan_kswapd_dma 1081421 1216199
pgscan_kswapd_dma32 58979118 46002810
pgscan_kswapd_normal 0 0
pgscan_kswapd_movable 0 0
pgscan_direct_dma 2015438 1086109
pgscan_direct_dma32 55787823 36101597
pgscan_direct_normal 0 0
pgscan_direct_movable 0 0
pginodesteal 3461 7281
slabs_scanned 564864 527616
kswapd_steal 2889797 1448082
kswapd_inodesteal 14827 14835
pageoutrun 43459 21562
allocstall 9653 4032
pgrotated 384216 228631
1.5) free numbers at the end of the tests
before patch:
total used free shared buffers cached
Mem: 474 467 7 0 0 236
-/+ buffers/cache: 230 243
Swap: 1023 418 605
after patch:
total used free shared buffers cached
Mem: 474 457 16 0 0 236
-/+ buffers/cache: 221 253
Swap: 1023 404 619
2) memory flushing in a file server
2.1) brief summary
The number of major faults from 50 to 3 during 10% cache hot reads.
That means this patch successfully stops major faults when the active file
list is slowly scanned when there are partially cache hot streaming IO.
2.2) test scenario
Do 100000 pread(size=110 pages, offset=(i*100) pages), where 10% of the
pages will be activated:
for i in `seq 0 100 10000000`; do echo $i 110; done > pattern-hot-10
iotrace.rb --load pattern-hot-10 --play /b/sparse
vmmon nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
and monitor /proc/vmstat during the time. The test box has 2G memory.
I carried out tests on fresh booted console as well as X desktop, and
fetched the vmstat numbers on
(1) begin: shortly after the big read IO starts;
(2) end: just before the big read IO stops;
(3) restore: the big read IO stops and the zsh working set restored
(4) restore X: after IO, switch back and forth between the urxvt and firefox
windows to restore their working set.
2.3) console mode results
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.29 VM_EXEC protection ON:
begin: 2481 2237 8694 630 0 574299
end: 275 231976 233914 633 776271 20933042
restore: 370 232154 234524 691 777183 20958453
2.6.29 VM_EXEC protection ON (second run):
begin: 2434 2237 8493 629 0 574195
end: 284 231970 233536 632 771918 20896129
restore: 399 232218 234789 690 774526 20957909
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 2479 2344 9659 210 0 579643
end: 284 232010 234142 260 772776 20917184
restore: 379 232159 234371 301 774888 20967849
The above console numbers show that
- The startup pgmajfault of 2.6.30-rc4-mm is merely 1/3 that of 2.6.29.
I'd attribute that improvement to the mmap readahead improvements :-)
- The pgmajfault increment during the file copy is 633-630=3 vs 260-210=50.
That's a huge improvement - which means with the VM_EXEC protection logic,
active mmap pages is pretty safe even under partially cache hot streaming IO.
- when active:inactive file lru size reaches 1:1, their scan rates is 1:20.8
under 10% cache hot IO. (computed with formula Dpgdeactivate:Dpgfree)
That roughly means the active mmap pages get 20.8 more chances to get
re-referenced to stay in memory.
- The absolute nr_mapped drops considerably to 1/9 during the big IO, and the
dropped pages are mostly inactive ones. The patch has almost no impact in
this aspect, that means it won't unnecessarily increase memory pressure.
(In contrast, your 20% mmap protection ratio will keep them all, and
therefore eliminate the extra 41 major faults to restore working set
of zsh etc.)
The iotrace.rb read throughput is
151.194384MB/s 284.198252s 100001x 450560b --load pattern-hot-10 --play /b/sparse
which means the inactive list is rotated at the speed of 250MB/s,
so a full scan of which takes about 3.5 seconds, while a full scan
of active file list takes about 77 seconds.
2.4) X mode results
We can reach roughly the same conclusions for X desktop:
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.30-rc4-mm VM_EXEC protection ON:
begin: 9740 8920 64075 561 0 678360
end: 768 218254 220029 565 798953 21057006
restore: 857 218543 220987 606 799462 21075710
restore X: 2414 218560 225344 797 799462 21080795
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 9368 5035 26389 554 0 633391
end: 770 218449 221230 661 646472 17832500
restore: 1113 218466 220978 710 649881 17905235
restore X: 2687 218650 225484 947 802700 21083584
- the absolute nr_mapped drops considerably (to 1/13 of the original size)
during the streaming IO.
- the delta of pgmajfault is 3 vs 107 during IO, or 236 vs 393
during the whole process.
Cc: Elladan <elladan@eskimo.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:12 +00:00
|
|
|
* so we ignore them here.
|
|
|
|
*/
|
2022-06-17 15:42:47 +00:00
|
|
|
if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
|
|
|
|
nr_rotated += folio_nr_pages(folio);
|
|
|
|
list_add(&folio->lru, &l_active);
|
vmscan: make mapped executable pages the first class citizen
Protect referenced PROT_EXEC mapped pages from being deactivated.
PROT_EXEC(or its internal presentation VM_EXEC) pages normally belong to some
currently running executables and their linked libraries, they shall really be
cached aggressively to provide good user experiences.
Thanks to Johannes Weiner for the advice to reuse the VMA walk in
page_referenced() to get the PROT_EXEC bit.
[more details]
( The consequences of this patch will have to be discussed together with
Rik van Riel's recent patch "vmscan: evict use-once pages first". )
( Some of the good points and insights are taken into this changelog.
Thanks to all the involved people for the great LKML discussions. )
the problem
===========
For a typical desktop, the most precious working set is composed of
*actively accessed*
(1) memory mapped executables
(2) and their anonymous pages
(3) and other files
(4) and the dcache/icache/.. slabs
while the least important data are
(5) infrequently used or use-once files
For a typical desktop, one major problem is busty and large amount of (5)
use-once files flushing out the working set.
Inside the working set, (4) dcache/icache have already been too sticky ;-)
So we only have to care (2) anonymous and (1)(3) file pages.
anonymous pages
===============
Anonymous pages are effectively immune to the streaming IO attack, because we
now have separate file/anon LRU lists. When the use-once files crowd into the
file LRU, the list's "quality" is significantly lowered. Therefore the scan
balance policy in get_scan_ratio() will choose to scan the (low quality) file
LRU much more frequently than the anon LRU.
file pages
==========
Rik proposed to *not* scan the active file LRU when the inactive list grows
larger than active list. This guarantees that when there are use-once streaming
IO, and the working set is not too large(so that active_size < inactive_size),
the active file LRU will *not* be scanned at all. So the not-too-large working
set can be well protected.
But there are also situations where the file working set is a bit large so that
(active_size >= inactive_size), or the streaming IOs are not purely use-once.
In these cases, the active list will be scanned slowly. Because the current
shrink_active_list() policy is to deactivate active pages regardless of their
referenced bits. The deactivated pages become susceptible to the streaming IO
attack: the inactive list could be scanned fast (500MB / 50MBps = 10s) so that
the deactivated pages don't have enough time to get re-referenced. Because a
user tend to switch between windows in intervals from seconds to minutes.
This patch holds mapped executable pages in the active list as long as they
are referenced during each full scan of the active list. Because the active
list is normally scanned much slower, they get longer grace time (eg. 100s)
for further references, which better matches the pace of user operations.
Therefore this patch greatly prolongs the in-cache time of executable code,
when there are moderate memory pressures.
before patch: guaranteed to be cached if reference intervals < I
after patch: guaranteed to be cached if reference intervals < I+A
(except when randomly reclaimed by the lumpy reclaim)
where
A = time to fully scan the active file LRU
I = time to fully scan the inactive file LRU
Note that normally A >> I.
side effects
============
This patch is safe in general, it restores the pre-2.6.28 mmap() behavior
but in a much smaller and well targeted scope.
One may worry about some one to abuse the PROT_EXEC heuristic. But as
Andrew Morton stated, there are other tricks to getting that sort of boost.
Another concern is the PROT_EXEC mapped pages growing large in rare cases,
and therefore hurting reclaim efficiency. But a sane application targeted for
large audience will never use PROT_EXEC for data mappings. If some home made
application tries to abuse that bit, it shall be aware of the consequences.
If it is abused to scale of 2/3 total memory, it gains nothing but overheads.
benchmarks
==========
1) memory tight desktop
1.1) brief summary
- clock time and major faults are reduced by 50%;
- pswpin numbers are reduced to ~1/3.
That means X desktop responsiveness is doubled under high memory/swap pressure.
1.2) test scenario
- nfsroot gnome desktop with 512M physical memory
- run some programs, and switch between the existing windows
after starting each new program.
1.3) progress timing (seconds)
before after programs
0.02 0.02 N xeyes
0.75 0.76 N firefox
2.02 1.88 N nautilus
3.36 3.17 N nautilus --browser
5.26 4.89 N gthumb
7.12 6.47 N gedit
9.22 8.16 N xpdf /usr/share/doc/shared-mime-info/shared-mime-info-spec.pdf
13.58 12.55 N xterm
15.87 14.57 N mlterm
18.63 17.06 N gnome-terminal
21.16 18.90 N urxvt
26.24 23.48 N gnome-system-monitor
28.72 26.52 N gnome-help
32.15 29.65 N gnome-dictionary
39.66 36.12 N /usr/games/sol
43.16 39.27 N /usr/games/gnometris
48.65 42.56 N /usr/games/gnect
53.31 47.03 N /usr/games/gtali
58.60 52.05 N /usr/games/iagno
65.77 55.42 N /usr/games/gnotravex
70.76 61.47 N /usr/games/mahjongg
76.15 67.11 N /usr/games/gnome-sudoku
86.32 75.15 N /usr/games/glines
92.21 79.70 N /usr/games/glchess
103.79 88.48 N /usr/games/gnomine
113.84 96.51 N /usr/games/gnotski
124.40 102.19 N /usr/games/gnibbles
137.41 114.93 N /usr/games/gnobots2
155.53 125.02 N /usr/games/blackjack
179.85 135.11 N /usr/games/same-gnome
224.49 154.50 N /usr/bin/gnome-window-properties
248.44 162.09 N /usr/bin/gnome-default-applications-properties
282.62 173.29 N /usr/bin/gnome-at-properties
323.72 188.21 N /usr/bin/gnome-typing-monitor
363.99 199.93 N /usr/bin/gnome-at-visual
394.21 206.95 N /usr/bin/gnome-sound-properties
435.14 224.49 N /usr/bin/gnome-at-mobility
463.05 234.11 N /usr/bin/gnome-keybinding-properties
503.75 248.59 N /usr/bin/gnome-about-me
554.00 276.27 N /usr/bin/gnome-display-properties
615.48 304.39 N /usr/bin/gnome-network-preferences
693.03 342.01 N /usr/bin/gnome-mouse-properties
759.90 388.58 N /usr/bin/gnome-appearance-properties
937.90 508.47 N /usr/bin/gnome-control-center
1109.75 587.57 N /usr/bin/gnome-keyboard-properties
1399.05 758.16 N : oocalc
1524.64 830.03 N : oodraw
1684.31 900.03 N : ooimpress
1874.04 993.91 N : oomath
2115.12 1081.89 N : ooweb
2369.02 1161.99 N : oowriter
Note that the last ": oo*" commands are actually commented out.
1.4) vmstat numbers (some relevant ones are marked with *)
before after
nr_free_pages 1293 3898
nr_inactive_anon 59956 53460
nr_active_anon 26815 30026
nr_inactive_file 2657 3218
nr_active_file 2019 2806
nr_unevictable 4 4
nr_mlock 4 4
nr_anon_pages 26706 27859
*nr_mapped 3542 4469
nr_file_pages 72232 67681
nr_dirty 1 0
nr_writeback 123 19
nr_slab_reclaimable 3375 3534
nr_slab_unreclaimable 11405 10665
nr_page_table_pages 8106 7864
nr_unstable 0 0
nr_bounce 0 0
*nr_vmscan_write 394776 230839
nr_writeback_temp 0 0
numa_hit 6843353 3318676
numa_miss 0 0
numa_foreign 0 0
numa_interleave 1719 1719
numa_local 6843353 3318676
numa_other 0 0
*pgpgin 5954683 2057175
*pgpgout 1578276 922744
*pswpin 1486615 512238
*pswpout 394568 230685
pgalloc_dma 277432 56602
pgalloc_dma32 6769477 3310348
pgalloc_normal 0 0
pgalloc_movable 0 0
pgfree 7048396 3371118
pgactivate 2036343 1471492
pgdeactivate 2189691 1612829
pgfault 3702176 3100702
*pgmajfault 452116 201343
pgrefill_dma 12185 7127
pgrefill_dma32 334384 653703
pgrefill_normal 0 0
pgrefill_movable 0 0
pgsteal_dma 74214 22179
pgsteal_dma32 3334164 1638029
pgsteal_normal 0 0
pgsteal_movable 0 0
pgscan_kswapd_dma 1081421 1216199
pgscan_kswapd_dma32 58979118 46002810
pgscan_kswapd_normal 0 0
pgscan_kswapd_movable 0 0
pgscan_direct_dma 2015438 1086109
pgscan_direct_dma32 55787823 36101597
pgscan_direct_normal 0 0
pgscan_direct_movable 0 0
pginodesteal 3461 7281
slabs_scanned 564864 527616
kswapd_steal 2889797 1448082
kswapd_inodesteal 14827 14835
pageoutrun 43459 21562
allocstall 9653 4032
pgrotated 384216 228631
1.5) free numbers at the end of the tests
before patch:
total used free shared buffers cached
Mem: 474 467 7 0 0 236
-/+ buffers/cache: 230 243
Swap: 1023 418 605
after patch:
total used free shared buffers cached
Mem: 474 457 16 0 0 236
-/+ buffers/cache: 221 253
Swap: 1023 404 619
2) memory flushing in a file server
2.1) brief summary
The number of major faults from 50 to 3 during 10% cache hot reads.
That means this patch successfully stops major faults when the active file
list is slowly scanned when there are partially cache hot streaming IO.
2.2) test scenario
Do 100000 pread(size=110 pages, offset=(i*100) pages), where 10% of the
pages will be activated:
for i in `seq 0 100 10000000`; do echo $i 110; done > pattern-hot-10
iotrace.rb --load pattern-hot-10 --play /b/sparse
vmmon nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
and monitor /proc/vmstat during the time. The test box has 2G memory.
I carried out tests on fresh booted console as well as X desktop, and
fetched the vmstat numbers on
(1) begin: shortly after the big read IO starts;
(2) end: just before the big read IO stops;
(3) restore: the big read IO stops and the zsh working set restored
(4) restore X: after IO, switch back and forth between the urxvt and firefox
windows to restore their working set.
2.3) console mode results
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.29 VM_EXEC protection ON:
begin: 2481 2237 8694 630 0 574299
end: 275 231976 233914 633 776271 20933042
restore: 370 232154 234524 691 777183 20958453
2.6.29 VM_EXEC protection ON (second run):
begin: 2434 2237 8493 629 0 574195
end: 284 231970 233536 632 771918 20896129
restore: 399 232218 234789 690 774526 20957909
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 2479 2344 9659 210 0 579643
end: 284 232010 234142 260 772776 20917184
restore: 379 232159 234371 301 774888 20967849
The above console numbers show that
- The startup pgmajfault of 2.6.30-rc4-mm is merely 1/3 that of 2.6.29.
I'd attribute that improvement to the mmap readahead improvements :-)
- The pgmajfault increment during the file copy is 633-630=3 vs 260-210=50.
That's a huge improvement - which means with the VM_EXEC protection logic,
active mmap pages is pretty safe even under partially cache hot streaming IO.
- when active:inactive file lru size reaches 1:1, their scan rates is 1:20.8
under 10% cache hot IO. (computed with formula Dpgdeactivate:Dpgfree)
That roughly means the active mmap pages get 20.8 more chances to get
re-referenced to stay in memory.
- The absolute nr_mapped drops considerably to 1/9 during the big IO, and the
dropped pages are mostly inactive ones. The patch has almost no impact in
this aspect, that means it won't unnecessarily increase memory pressure.
(In contrast, your 20% mmap protection ratio will keep them all, and
therefore eliminate the extra 41 major faults to restore working set
of zsh etc.)
The iotrace.rb read throughput is
151.194384MB/s 284.198252s 100001x 450560b --load pattern-hot-10 --play /b/sparse
which means the inactive list is rotated at the speed of 250MB/s,
so a full scan of which takes about 3.5 seconds, while a full scan
of active file list takes about 77 seconds.
2.4) X mode results
We can reach roughly the same conclusions for X desktop:
nr_mapped nr_active_file nr_inactive_file pgmajfault pgdeactivate pgfree
2.6.30-rc4-mm VM_EXEC protection ON:
begin: 9740 8920 64075 561 0 678360
end: 768 218254 220029 565 798953 21057006
restore: 857 218543 220987 606 799462 21075710
restore X: 2414 218560 225344 797 799462 21080795
2.6.30-rc4-mm VM_EXEC protection OFF:
begin: 9368 5035 26389 554 0 633391
end: 770 218449 221230 661 646472 17832500
restore: 1113 218466 220978 710 649881 17905235
restore X: 2687 218650 225484 947 802700 21083584
- the absolute nr_mapped drops considerably (to 1/13 of the original size)
during the streaming IO.
- the delta of pgmajfault is 3 vs 107 during IO, or 236 vs 393
during the whole process.
Cc: Elladan <elladan@eskimo.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:12 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
2008-10-19 03:26:35 +00:00
|
|
|
|
2022-06-17 15:42:47 +00:00
|
|
|
folio_clear_active(folio); /* we are de-activating */
|
|
|
|
folio_set_workingset(folio);
|
|
|
|
list_add(&folio->lru, &l_inactive);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2009-01-06 22:40:13 +00:00
|
|
|
/*
|
2022-06-17 15:42:47 +00:00
|
|
|
* Move folios back to the lru list.
|
2009-01-06 22:40:13 +00:00
|
|
|
*/
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
2008-10-19 03:26:34 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_activate = move_folios_to_lru(lruvec, &l_active);
|
|
|
|
nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
|
2022-06-17 15:42:47 +00:00
|
|
|
/* Keep all free folios in l_active list */
|
2019-05-14 00:16:57 +00:00
|
|
|
list_splice(&l_inactive, &l_active);
|
2019-05-14 00:16:54 +00:00
|
|
|
|
|
|
|
__count_vm_events(PGDEACTIVATE, nr_deactivate);
|
|
|
|
__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
|
|
|
|
|
2016-07-28 22:45:31 +00:00
|
|
|
__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
|
2020-12-15 20:34:29 +00:00
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
2012-01-13 01:19:56 +00:00
|
|
|
|
mm: vmscan: make rotations a secondary factor in balancing anon vs file
We noticed a 2% webserver throughput regression after upgrading from 5.6.
This could be tracked down to a shift in the anon/file reclaim balance
(confirmed with swappiness) that resulted in worse reclaim efficiency and
thus more kswapd activity for the same outcome.
The change that exposed the problem is aae466b0052e ("mm/swap: implement
workingset detection for anonymous LRU"). By qualifying swapins based on
their refault distance, it lowered the cost of anon reclaim in this
workload, in turn causing (much) more anon scanning than before. Scanning
the anon list is more expensive due to the higher ratio of mmapped pages
that may rotate during reclaim, and so the result was an increase in %sys
time.
Right now, rotations aren't considered a cost when balancing scan pressure
between LRUs. We can end up with very few file refaults putting all the
scan pressure on hot anon pages that are rotated en masse, don't get
reclaimed, and never push back on the file LRU again. We still only
reclaim file cache in that case, but we burn a lot CPU rotating anon
pages. It's "fair" from an LRU age POV, but doesn't reflect the real cost
it imposes on the system.
Consider rotations as a secondary factor in balancing the LRUs. This
doesn't attempt to make a precise comparison between IO cost and CPU cost,
it just says: if reloads are about comparable between the lists, or
rotations are overwhelmingly different, adjust for CPU work.
This fixed the regression on our webservers. It has since been deployed
to the entire Meta fleet and hasn't caused any problems.
Link: https://lkml.kernel.org/r/20221013193113.726425-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Rik van Riel <riel@surriel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-10-13 19:31:13 +00:00
|
|
|
if (nr_rotated)
|
|
|
|
lru_note_cost(lruvec, file, 0, nr_rotated);
|
2019-05-14 00:16:57 +00:00
|
|
|
mem_cgroup_uncharge_list(&l_active);
|
|
|
|
free_unref_page_list(&l_active);
|
2017-02-22 23:44:18 +00:00
|
|
|
trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
|
|
|
|
nr_deactivate, nr_rotated, sc->priority, file);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
static unsigned int reclaim_folio_list(struct list_head *folio_list,
|
2022-05-13 03:22:59 +00:00
|
|
|
struct pglist_data *pgdat)
|
2019-09-25 23:49:15 +00:00
|
|
|
{
|
|
|
|
struct reclaim_stat dummy_stat;
|
2022-05-13 03:22:59 +00:00
|
|
|
unsigned int nr_reclaimed;
|
|
|
|
struct folio *folio;
|
2019-09-25 23:49:15 +00:00
|
|
|
struct scan_control sc = {
|
|
|
|
.gfp_mask = GFP_KERNEL,
|
|
|
|
.may_writepage = 1,
|
|
|
|
.may_unmap = 1,
|
|
|
|
.may_swap = 1,
|
2021-09-02 21:59:16 +00:00
|
|
|
.no_demotion = 1,
|
2019-09-25 23:49:15 +00:00
|
|
|
};
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false);
|
|
|
|
while (!list_empty(folio_list)) {
|
|
|
|
folio = lru_to_folio(folio_list);
|
2022-05-13 03:22:59 +00:00
|
|
|
list_del(&folio->lru);
|
|
|
|
folio_putback_lru(folio);
|
|
|
|
}
|
|
|
|
|
|
|
|
return nr_reclaimed;
|
|
|
|
}
|
|
|
|
|
2022-06-17 15:42:48 +00:00
|
|
|
unsigned long reclaim_pages(struct list_head *folio_list)
|
2022-05-13 03:22:59 +00:00
|
|
|
{
|
2022-05-13 03:23:00 +00:00
|
|
|
int nid;
|
2022-05-13 03:22:59 +00:00
|
|
|
unsigned int nr_reclaimed = 0;
|
2022-06-17 15:42:48 +00:00
|
|
|
LIST_HEAD(node_folio_list);
|
2022-05-13 03:22:59 +00:00
|
|
|
unsigned int noreclaim_flag;
|
|
|
|
|
2022-06-17 15:42:48 +00:00
|
|
|
if (list_empty(folio_list))
|
2022-05-13 03:23:00 +00:00
|
|
|
return nr_reclaimed;
|
|
|
|
|
2021-07-01 01:49:48 +00:00
|
|
|
noreclaim_flag = memalloc_noreclaim_save();
|
|
|
|
|
2022-06-17 15:42:48 +00:00
|
|
|
nid = folio_nid(lru_to_folio(folio_list));
|
2022-05-13 03:23:00 +00:00
|
|
|
do {
|
2022-06-17 15:42:48 +00:00
|
|
|
struct folio *folio = lru_to_folio(folio_list);
|
2019-09-25 23:49:15 +00:00
|
|
|
|
2022-06-17 15:42:48 +00:00
|
|
|
if (nid == folio_nid(folio)) {
|
|
|
|
folio_clear_active(folio);
|
|
|
|
list_move(&folio->lru, &node_folio_list);
|
2019-09-25 23:49:15 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
|
2022-06-17 15:42:48 +00:00
|
|
|
nid = folio_nid(lru_to_folio(folio_list));
|
|
|
|
} while (!list_empty(folio_list));
|
2019-09-25 23:49:15 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
|
2019-09-25 23:49:15 +00:00
|
|
|
|
2021-07-01 01:49:48 +00:00
|
|
|
memalloc_noreclaim_restore(noreclaim_flag);
|
|
|
|
|
2019-09-25 23:49:15 +00:00
|
|
|
return nr_reclaimed;
|
|
|
|
}
|
|
|
|
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
|
|
|
|
struct lruvec *lruvec, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
if (is_active_lru(lru)) {
|
|
|
|
if (sc->may_deactivate & (1 << is_file_lru(lru)))
|
|
|
|
shrink_active_list(nr_to_scan, lruvec, sc, lru);
|
|
|
|
else
|
|
|
|
sc->skipped_deactivate = 1;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
|
|
|
|
}
|
|
|
|
|
2016-05-20 23:56:31 +00:00
|
|
|
/*
|
|
|
|
* The inactive anon list should be small enough that the VM never has
|
|
|
|
* to do too much work.
|
2009-01-08 02:08:18 +00:00
|
|
|
*
|
2016-05-20 23:56:31 +00:00
|
|
|
* The inactive file list should be small enough to leave most memory
|
|
|
|
* to the established workingset on the scan-resistant active list,
|
|
|
|
* but large enough to avoid thrashing the aggregate readahead window.
|
2009-06-16 22:32:28 +00:00
|
|
|
*
|
2016-05-20 23:56:31 +00:00
|
|
|
* Both inactive lists should also be large enough that each inactive
|
2022-09-02 19:45:57 +00:00
|
|
|
* folio has a chance to be referenced again before it is reclaimed.
|
2009-06-16 22:32:28 +00:00
|
|
|
*
|
2017-05-03 21:55:03 +00:00
|
|
|
* If that fails and refaulting is observed, the inactive list grows.
|
|
|
|
*
|
2022-09-02 19:45:57 +00:00
|
|
|
* The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
|
2017-11-16 01:34:15 +00:00
|
|
|
* on this LRU, maintained by the pageout code. An inactive_ratio
|
2022-09-02 19:45:57 +00:00
|
|
|
* of 3 means 3:1 or 25% of the folios are kept on the inactive list.
|
2009-06-16 22:32:28 +00:00
|
|
|
*
|
2016-05-20 23:56:31 +00:00
|
|
|
* total target max
|
|
|
|
* memory ratio inactive
|
|
|
|
* -------------------------------------
|
|
|
|
* 10MB 1 5MB
|
|
|
|
* 100MB 1 50MB
|
|
|
|
* 1GB 3 250MB
|
|
|
|
* 10GB 10 0.9GB
|
|
|
|
* 100GB 31 3GB
|
|
|
|
* 1TB 101 10GB
|
|
|
|
* 10TB 320 32GB
|
2009-06-16 22:32:28 +00:00
|
|
|
*/
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
|
2009-06-16 22:32:28 +00:00
|
|
|
{
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
|
2017-05-03 21:55:03 +00:00
|
|
|
unsigned long inactive, active;
|
|
|
|
unsigned long inactive_ratio;
|
2016-05-20 23:56:31 +00:00
|
|
|
unsigned long gb;
|
2013-02-23 00:35:19 +00:00
|
|
|
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
|
|
|
|
active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
|
2016-07-28 22:47:34 +00:00
|
|
|
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
gb = (inactive + active) >> (30 - PAGE_SHIFT);
|
2020-08-12 01:30:54 +00:00
|
|
|
if (gb)
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
inactive_ratio = int_sqrt(10 * gb);
|
|
|
|
else
|
|
|
|
inactive_ratio = 1;
|
2017-02-22 23:45:58 +00:00
|
|
|
|
2016-05-20 23:56:31 +00:00
|
|
|
return inactive * inactive_ratio < active;
|
2009-12-15 01:59:48 +00:00
|
|
|
}
|
|
|
|
|
2013-02-23 00:32:17 +00:00
|
|
|
enum scan_balance {
|
|
|
|
SCAN_EQUAL,
|
|
|
|
SCAN_FRACT,
|
|
|
|
SCAN_ANON,
|
|
|
|
SCAN_FILE,
|
|
|
|
};
|
|
|
|
|
2022-09-18 08:00:00 +00:00
|
|
|
static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
unsigned long file;
|
|
|
|
struct lruvec *target_lruvec;
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
if (lru_gen_enabled())
|
|
|
|
return;
|
|
|
|
|
2022-09-18 08:00:00 +00:00
|
|
|
target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Flush the memory cgroup stats, so that we read accurate per-memcg
|
|
|
|
* lruvec stats for heuristics.
|
|
|
|
*/
|
|
|
|
mem_cgroup_flush_stats();
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Determine the scan balance between anon and file LRUs.
|
|
|
|
*/
|
|
|
|
spin_lock_irq(&target_lruvec->lru_lock);
|
|
|
|
sc->anon_cost = target_lruvec->anon_cost;
|
|
|
|
sc->file_cost = target_lruvec->file_cost;
|
|
|
|
spin_unlock_irq(&target_lruvec->lru_lock);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Target desirable inactive:active list ratios for the anon
|
|
|
|
* and file LRU lists.
|
|
|
|
*/
|
|
|
|
if (!sc->force_deactivate) {
|
|
|
|
unsigned long refaults;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* When refaults are being observed, it means a new
|
|
|
|
* workingset is being established. Deactivate to get
|
|
|
|
* rid of any stale active pages quickly.
|
|
|
|
*/
|
|
|
|
refaults = lruvec_page_state(target_lruvec,
|
|
|
|
WORKINGSET_ACTIVATE_ANON);
|
|
|
|
if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
|
|
|
|
inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
|
|
|
|
sc->may_deactivate |= DEACTIVATE_ANON;
|
|
|
|
else
|
|
|
|
sc->may_deactivate &= ~DEACTIVATE_ANON;
|
|
|
|
|
|
|
|
refaults = lruvec_page_state(target_lruvec,
|
|
|
|
WORKINGSET_ACTIVATE_FILE);
|
|
|
|
if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
|
|
|
|
inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
|
|
|
|
sc->may_deactivate |= DEACTIVATE_FILE;
|
|
|
|
else
|
|
|
|
sc->may_deactivate &= ~DEACTIVATE_FILE;
|
|
|
|
} else
|
|
|
|
sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we have plenty of inactive file pages that aren't
|
|
|
|
* thrashing, try to reclaim those first before touching
|
|
|
|
* anonymous pages.
|
|
|
|
*/
|
|
|
|
file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
|
|
|
|
if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
|
|
|
|
sc->cache_trim_mode = 1;
|
|
|
|
else
|
|
|
|
sc->cache_trim_mode = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Prevent the reclaimer from falling into the cache trap: as
|
|
|
|
* cache pages start out inactive, every cache fault will tip
|
|
|
|
* the scan balance towards the file LRU. And as the file LRU
|
|
|
|
* shrinks, so does the window for rotation from references.
|
|
|
|
* This means we have a runaway feedback loop where a tiny
|
|
|
|
* thrashing file LRU becomes infinitely more attractive than
|
|
|
|
* anon pages. Try to detect this based on file LRU size.
|
|
|
|
*/
|
|
|
|
if (!cgroup_reclaim(sc)) {
|
|
|
|
unsigned long total_high_wmark = 0;
|
|
|
|
unsigned long free, anon;
|
|
|
|
int z;
|
|
|
|
|
|
|
|
free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
|
|
|
|
file = node_page_state(pgdat, NR_ACTIVE_FILE) +
|
|
|
|
node_page_state(pgdat, NR_INACTIVE_FILE);
|
|
|
|
|
|
|
|
for (z = 0; z < MAX_NR_ZONES; z++) {
|
|
|
|
struct zone *zone = &pgdat->node_zones[z];
|
|
|
|
|
|
|
|
if (!managed_zone(zone))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
total_high_wmark += high_wmark_pages(zone);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Consider anon: if that's low too, this isn't a
|
|
|
|
* runaway file reclaim problem, but rather just
|
|
|
|
* extreme pressure. Reclaim as per usual then.
|
|
|
|
*/
|
|
|
|
anon = node_page_state(pgdat, NR_INACTIVE_ANON);
|
|
|
|
|
|
|
|
sc->file_is_tiny =
|
|
|
|
file + free <= total_high_wmark &&
|
|
|
|
!(sc->may_deactivate & DEACTIVATE_ANON) &&
|
|
|
|
anon >> sc->priority;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-10-19 03:26:32 +00:00
|
|
|
/*
|
|
|
|
* Determine how aggressively the anon and file LRU lists should be
|
2022-04-29 06:16:04 +00:00
|
|
|
* scanned.
|
2008-10-19 03:26:32 +00:00
|
|
|
*
|
2022-09-02 19:45:57 +00:00
|
|
|
* nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
|
|
|
|
* nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
|
2008-10-19 03:26:32 +00:00
|
|
|
*/
|
2019-12-01 01:55:46 +00:00
|
|
|
static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
|
|
|
|
unsigned long *nr)
|
2008-10-19 03:26:32 +00:00
|
|
|
{
|
2021-09-02 21:59:26 +00:00
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(lruvec);
|
2019-12-01 01:55:46 +00:00
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
2020-06-03 23:03:13 +00:00
|
|
|
unsigned long anon_cost, file_cost, total_cost;
|
2016-01-20 23:02:59 +00:00
|
|
|
int swappiness = mem_cgroup_swappiness(memcg);
|
2020-10-16 03:09:55 +00:00
|
|
|
u64 fraction[ANON_AND_FILE];
|
2013-02-23 00:32:17 +00:00
|
|
|
u64 denominator = 0; /* gcc */
|
|
|
|
enum scan_balance scan_balance;
|
2008-10-19 03:26:32 +00:00
|
|
|
unsigned long ap, fp;
|
2012-01-13 01:20:01 +00:00
|
|
|
enum lru_list lru;
|
vmscan: prevent get_scan_ratio() rounding errors
get_scan_ratio() calculates percentage and if the percentage is < 1%, it
will round percentage down to 0% and cause we completely ignore scanning
anon/file pages to reclaim memory even the total anon/file pages are very
big.
To avoid underflow, we don't use percentage, instead we directly calculate
how many pages should be scaned. In this way, we should get several
scanned pages for < 1% percent.
This has some benefits:
1. increase our calculation precision
2. making our scan more smoothly. Without this, if percent[x] is
underflow, shrink_zone() doesn't scan any pages and suddenly it scans
all pages when priority is zero. With this, even priority isn't zero,
shrink_zone() gets chance to scan some pages.
Note, this patch doesn't really change logics, but just increase
precision. For system with a lot of memory, this might slightly changes
behavior. For example, in a sequential file read workload, without the
patch, we don't swap any anon pages. With it, if anon memory size is
bigger than 16G, we will see one anon page swapped. The 16G is calculated
as PAGE_SIZE * priority(4096) * (fp/ap). fp/ap is assumed to be 1024
which is common in this workload. So the impact sounds not a big deal.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:36 +00:00
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
/* If we have no swap space, do not bother scanning anon folios. */
|
2021-09-02 21:59:26 +00:00
|
|
|
if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
|
2013-02-23 00:32:17 +00:00
|
|
|
scan_balance = SCAN_FILE;
|
vmscan: prevent get_scan_ratio() rounding errors
get_scan_ratio() calculates percentage and if the percentage is < 1%, it
will round percentage down to 0% and cause we completely ignore scanning
anon/file pages to reclaim memory even the total anon/file pages are very
big.
To avoid underflow, we don't use percentage, instead we directly calculate
how many pages should be scaned. In this way, we should get several
scanned pages for < 1% percent.
This has some benefits:
1. increase our calculation precision
2. making our scan more smoothly. Without this, if percent[x] is
underflow, shrink_zone() doesn't scan any pages and suddenly it scans
all pages when priority is zero. With this, even priority isn't zero,
shrink_zone() gets chance to scan some pages.
Note, this patch doesn't really change logics, but just increase
precision. For system with a lot of memory, this might slightly changes
behavior. For example, in a sequential file read workload, without the
patch, we don't swap any anon pages. With it, if anon memory size is
bigger than 16G, we will see one anon page swapped. The 16G is calculated
as PAGE_SIZE * priority(4096) * (fp/ap). fp/ap is assumed to be 1024
which is common in this workload. So the impact sounds not a big deal.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:36 +00:00
|
|
|
goto out;
|
|
|
|
}
|
2008-10-19 03:26:32 +00:00
|
|
|
|
2013-02-23 00:32:14 +00:00
|
|
|
/*
|
|
|
|
* Global reclaim will swap to prevent OOM even with no
|
|
|
|
* swappiness, but memcg users want to use this knob to
|
|
|
|
* disable swapping for individual groups completely when
|
|
|
|
* using the memory controller's swap limit feature would be
|
|
|
|
* too expensive.
|
|
|
|
*/
|
2019-12-01 01:55:40 +00:00
|
|
|
if (cgroup_reclaim(sc) && !swappiness) {
|
2013-02-23 00:32:17 +00:00
|
|
|
scan_balance = SCAN_FILE;
|
2013-02-23 00:32:14 +00:00
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Do not apply any pressure balancing cleverness when the
|
|
|
|
* system is close to OOM, scan both anon and file equally
|
|
|
|
* (unless the swappiness setting disagrees with swapping).
|
|
|
|
*/
|
2014-08-06 23:06:17 +00:00
|
|
|
if (!sc->priority && swappiness) {
|
2013-02-23 00:32:17 +00:00
|
|
|
scan_balance = SCAN_EQUAL;
|
2013-02-23 00:32:14 +00:00
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
2014-05-06 19:50:07 +00:00
|
|
|
/*
|
mm: vmscan: move file exhaustion detection to the node level
Patch series "mm: fix page aging across multiple cgroups".
When applications are put into unconfigured cgroups for memory accounting
purposes, the cgrouping itself should not change the behavior of the page
reclaim code. We expect the VM to reclaim the coldest pages in the
system. But right now the VM can reclaim hot pages in one cgroup while
there is eligible cold cache in others.
This is because one part of the reclaim algorithm isn't truly cgroup
hierarchy aware: the inactive/active list balancing. That is the part
that is supposed to protect hot cache data from one-off streaming IO.
The recursive cgroup reclaim scheme will scan and rotate the physical LRU
lists of each eligible cgroup at the same rate in a round-robin fashion,
thereby establishing a relative order among the pages of all those
cgroups. However, the inactive/active balancing decisions are made
locally within each cgroup, so when a cgroup is running low on cold pages,
its hot pages will get reclaimed - even when sibling cgroups have plenty
of cold cache eligible in the same reclaim run.
For example:
[root@ham ~]# head -n1 /proc/meminfo
MemTotal: 1016336 kB
[root@ham ~]# ./reclaimtest2.sh
Establishing 50M active files in cgroup A...
Hot pages cached: 12800/12800 workingset-a
Linearly scanning through 18G of file data in cgroup B:
real 0m4.269s
user 0m0.051s
sys 0m4.182s
Hot pages cached: 134/12800 workingset-a
The streaming IO in B, which doesn't benefit from caching at all, pushes
out most of the workingset in A.
Solution
This series fixes the problem by elevating inactive/active balancing
decisions to the toplevel of the reclaim run. This is either a cgroup
that hit its limit, or straight-up global reclaim if there is physical
memory pressure. From there, it takes a recursive view of the cgroup
subtree to decide whether page deactivation is necessary.
In the test above, the VM will then recognize that cgroup B has plenty of
eligible cold cache, and that the hot pages in A can be spared:
[root@ham ~]# ./reclaimtest2.sh
Establishing 50M active files in cgroup A...
Hot pages cached: 12800/12800 workingset-a
Linearly scanning through 18G of file data in cgroup B:
real 0m4.244s
user 0m0.064s
sys 0m4.177s
Hot pages cached: 12800/12800 workingset-a
Implementation
Whether active pages can be deactivated or not is influenced by two
factors: the inactive list dropping below a minimum size relative to the
active list, and the occurence of refaults.
This patch series first moves refault detection to the reclaim root, then
enforces the minimum inactive size based on a recursive view of the cgroup
tree's LRUs.
History
Note that this actually never worked correctly in Linux cgroups. In the
past it worked for global reclaim and leaf limit reclaim only (we used to
have two physical LRU linkages per page), but it never worked for
intermediate limit reclaim over multiple leaf cgroups.
We're noticing this now because 1) we're putting everything into cgroups
for accounting, not just the things we want to control and 2) we're moving
away from leaf limits that invoke reclaim on individual cgroups, toward
large tree reclaim, triggered by high-level limits, or physical memory
pressure that is influenced by local protections such as memory.low and
memory.min instead.
This patch (of 3):
When file pages are lower than the watermark on a node, we try to force
scan anonymous pages to counter-act the balancing algorithms preference
for new file pages when they are likely thrashing. This is a node-level
decision, but it's currently made each time we look at an lruvec. This is
unnecessarily expensive and also a layering violation that makes the code
harder to understand.
Clean this up by making the check once per node and setting a flag in the
scan_control.
Link: http://lkml.kernel.org/r/20191107205334.158354-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:56 +00:00
|
|
|
* If the system is almost out of file pages, force-scan anon.
|
2014-05-06 19:50:07 +00:00
|
|
|
*/
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
if (sc->file_is_tiny) {
|
mm: vmscan: move file exhaustion detection to the node level
Patch series "mm: fix page aging across multiple cgroups".
When applications are put into unconfigured cgroups for memory accounting
purposes, the cgrouping itself should not change the behavior of the page
reclaim code. We expect the VM to reclaim the coldest pages in the
system. But right now the VM can reclaim hot pages in one cgroup while
there is eligible cold cache in others.
This is because one part of the reclaim algorithm isn't truly cgroup
hierarchy aware: the inactive/active list balancing. That is the part
that is supposed to protect hot cache data from one-off streaming IO.
The recursive cgroup reclaim scheme will scan and rotate the physical LRU
lists of each eligible cgroup at the same rate in a round-robin fashion,
thereby establishing a relative order among the pages of all those
cgroups. However, the inactive/active balancing decisions are made
locally within each cgroup, so when a cgroup is running low on cold pages,
its hot pages will get reclaimed - even when sibling cgroups have plenty
of cold cache eligible in the same reclaim run.
For example:
[root@ham ~]# head -n1 /proc/meminfo
MemTotal: 1016336 kB
[root@ham ~]# ./reclaimtest2.sh
Establishing 50M active files in cgroup A...
Hot pages cached: 12800/12800 workingset-a
Linearly scanning through 18G of file data in cgroup B:
real 0m4.269s
user 0m0.051s
sys 0m4.182s
Hot pages cached: 134/12800 workingset-a
The streaming IO in B, which doesn't benefit from caching at all, pushes
out most of the workingset in A.
Solution
This series fixes the problem by elevating inactive/active balancing
decisions to the toplevel of the reclaim run. This is either a cgroup
that hit its limit, or straight-up global reclaim if there is physical
memory pressure. From there, it takes a recursive view of the cgroup
subtree to decide whether page deactivation is necessary.
In the test above, the VM will then recognize that cgroup B has plenty of
eligible cold cache, and that the hot pages in A can be spared:
[root@ham ~]# ./reclaimtest2.sh
Establishing 50M active files in cgroup A...
Hot pages cached: 12800/12800 workingset-a
Linearly scanning through 18G of file data in cgroup B:
real 0m4.244s
user 0m0.064s
sys 0m4.177s
Hot pages cached: 12800/12800 workingset-a
Implementation
Whether active pages can be deactivated or not is influenced by two
factors: the inactive list dropping below a minimum size relative to the
active list, and the occurence of refaults.
This patch series first moves refault detection to the reclaim root, then
enforces the minimum inactive size based on a recursive view of the cgroup
tree's LRUs.
History
Note that this actually never worked correctly in Linux cgroups. In the
past it worked for global reclaim and leaf limit reclaim only (we used to
have two physical LRU linkages per page), but it never worked for
intermediate limit reclaim over multiple leaf cgroups.
We're noticing this now because 1) we're putting everything into cgroups
for accounting, not just the things we want to control and 2) we're moving
away from leaf limits that invoke reclaim on individual cgroups, toward
large tree reclaim, triggered by high-level limits, or physical memory
pressure that is influenced by local protections such as memory.low and
memory.min instead.
This patch (of 3):
When file pages are lower than the watermark on a node, we try to force
scan anonymous pages to counter-act the balancing algorithms preference
for new file pages when they are likely thrashing. This is a node-level
decision, but it's currently made each time we look at an lruvec. This is
unnecessarily expensive and also a layering violation that makes the code
harder to understand.
Clean this up by making the check once per node and setting a flag in the
scan_control.
Link: http://lkml.kernel.org/r/20191107205334.158354-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:56 +00:00
|
|
|
scan_balance = SCAN_ANON;
|
|
|
|
goto out;
|
2014-05-06 19:50:07 +00:00
|
|
|
}
|
|
|
|
|
2013-02-23 00:32:10 +00:00
|
|
|
/*
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
* If there is enough inactive page cache, we do not reclaim
|
|
|
|
* anything from the anonymous working right now.
|
2013-02-23 00:32:10 +00:00
|
|
|
*/
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
if (sc->cache_trim_mode) {
|
2013-02-23 00:32:17 +00:00
|
|
|
scan_balance = SCAN_FILE;
|
2013-02-23 00:32:10 +00:00
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
2013-02-23 00:32:17 +00:00
|
|
|
scan_balance = SCAN_FRACT;
|
2010-08-10 00:19:51 +00:00
|
|
|
/*
|
mm: balance LRU lists based on relative thrashing
Since the LRUs were split into anon and file lists, the VM has been
balancing between page cache and anonymous pages based on per-list ratios
of scanned vs. rotated pages. In most cases that tips page reclaim
towards the list that is easier to reclaim and has the fewest actively
used pages, but there are a few problems with it:
1. Refaults and LRU rotations are weighted the same way, even though
one costs IO and the other costs a bit of CPU.
2. The less we scan an LRU list based on already observed rotations,
the more we increase the sampling interval for new references, and
rotations become even more likely on that list. This can enter a
death spiral in which we stop looking at one list completely until
the other one is all but annihilated by page reclaim.
Since commit a528910e12ec ("mm: thrash detection-based file cache sizing")
we have refault detection for the page cache. Along with swapin events,
they are good indicators of when the file or anon list, respectively, is
too small for its workingset and needs to grow.
For example, if the page cache is thrashing, the cache pages need more
time in memory, while there may be colder pages on the anonymous list.
Likewise, if swapped pages are faulting back in, it indicates that we
reclaim anonymous pages too aggressively and should back off.
Replace LRU rotations with refaults and swapins as the basis for relative
reclaim cost of the two LRUs. This will have the VM target list balances
that incur the least amount of IO on aggregate.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@surriel.com>
Link: http://lkml.kernel.org/r/20200520232525.798933-12-hannes@cmpxchg.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:03:03 +00:00
|
|
|
* Calculate the pressure balance between anon and file pages.
|
|
|
|
*
|
|
|
|
* The amount of pressure we put on each LRU is inversely
|
|
|
|
* proportional to the cost of reclaiming each list, as
|
|
|
|
* determined by the share of pages that are refaulting, times
|
|
|
|
* the relative IO cost of bringing back a swapped out
|
|
|
|
* anonymous page vs reloading a filesystem page (swappiness).
|
|
|
|
*
|
2020-06-03 23:03:13 +00:00
|
|
|
* Although we limit that influence to ensure no list gets
|
|
|
|
* left behind completely: at least a third of the pressure is
|
|
|
|
* applied, before swappiness.
|
|
|
|
*
|
mm: balance LRU lists based on relative thrashing
Since the LRUs were split into anon and file lists, the VM has been
balancing between page cache and anonymous pages based on per-list ratios
of scanned vs. rotated pages. In most cases that tips page reclaim
towards the list that is easier to reclaim and has the fewest actively
used pages, but there are a few problems with it:
1. Refaults and LRU rotations are weighted the same way, even though
one costs IO and the other costs a bit of CPU.
2. The less we scan an LRU list based on already observed rotations,
the more we increase the sampling interval for new references, and
rotations become even more likely on that list. This can enter a
death spiral in which we stop looking at one list completely until
the other one is all but annihilated by page reclaim.
Since commit a528910e12ec ("mm: thrash detection-based file cache sizing")
we have refault detection for the page cache. Along with swapin events,
they are good indicators of when the file or anon list, respectively, is
too small for its workingset and needs to grow.
For example, if the page cache is thrashing, the cache pages need more
time in memory, while there may be colder pages on the anonymous list.
Likewise, if swapped pages are faulting back in, it indicates that we
reclaim anonymous pages too aggressively and should back off.
Replace LRU rotations with refaults and swapins as the basis for relative
reclaim cost of the two LRUs. This will have the VM target list balances
that incur the least amount of IO on aggregate.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@surriel.com>
Link: http://lkml.kernel.org/r/20200520232525.798933-12-hannes@cmpxchg.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 23:03:03 +00:00
|
|
|
* With swappiness at 100, anon and file have equal IO cost.
|
2010-08-10 00:19:51 +00:00
|
|
|
*/
|
2020-06-03 23:03:13 +00:00
|
|
|
total_cost = sc->anon_cost + sc->file_cost;
|
|
|
|
anon_cost = total_cost + sc->anon_cost;
|
|
|
|
file_cost = total_cost + sc->file_cost;
|
|
|
|
total_cost = anon_cost + file_cost;
|
2010-08-10 00:19:51 +00:00
|
|
|
|
2020-06-03 23:03:13 +00:00
|
|
|
ap = swappiness * (total_cost + 1);
|
|
|
|
ap /= anon_cost + 1;
|
2008-10-19 03:26:32 +00:00
|
|
|
|
2020-06-03 23:03:13 +00:00
|
|
|
fp = (200 - swappiness) * (total_cost + 1);
|
|
|
|
fp /= file_cost + 1;
|
2008-10-19 03:26:32 +00:00
|
|
|
|
vmscan: prevent get_scan_ratio() rounding errors
get_scan_ratio() calculates percentage and if the percentage is < 1%, it
will round percentage down to 0% and cause we completely ignore scanning
anon/file pages to reclaim memory even the total anon/file pages are very
big.
To avoid underflow, we don't use percentage, instead we directly calculate
how many pages should be scaned. In this way, we should get several
scanned pages for < 1% percent.
This has some benefits:
1. increase our calculation precision
2. making our scan more smoothly. Without this, if percent[x] is
underflow, shrink_zone() doesn't scan any pages and suddenly it scans
all pages when priority is zero. With this, even priority isn't zero,
shrink_zone() gets chance to scan some pages.
Note, this patch doesn't really change logics, but just increase
precision. For system with a lot of memory, this might slightly changes
behavior. For example, in a sequential file read workload, without the
patch, we don't swap any anon pages. With it, if anon memory size is
bigger than 16G, we will see one anon page swapped. The 16G is calculated
as PAGE_SIZE * priority(4096) * (fp/ap). fp/ap is assumed to be 1024
which is common in this workload. So the impact sounds not a big deal.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:36 +00:00
|
|
|
fraction[0] = ap;
|
|
|
|
fraction[1] = fp;
|
2020-06-03 23:02:50 +00:00
|
|
|
denominator = ap + fp;
|
vmscan: prevent get_scan_ratio() rounding errors
get_scan_ratio() calculates percentage and if the percentage is < 1%, it
will round percentage down to 0% and cause we completely ignore scanning
anon/file pages to reclaim memory even the total anon/file pages are very
big.
To avoid underflow, we don't use percentage, instead we directly calculate
how many pages should be scaned. In this way, we should get several
scanned pages for < 1% percent.
This has some benefits:
1. increase our calculation precision
2. making our scan more smoothly. Without this, if percent[x] is
underflow, shrink_zone() doesn't scan any pages and suddenly it scans
all pages when priority is zero. With this, even priority isn't zero,
shrink_zone() gets chance to scan some pages.
Note, this patch doesn't really change logics, but just increase
precision. For system with a lot of memory, this might slightly changes
behavior. For example, in a sequential file read workload, without the
patch, we don't swap any anon pages. With it, if anon memory size is
bigger than 16G, we will see one anon page swapped. The 16G is calculated
as PAGE_SIZE * priority(4096) * (fp/ap). fp/ap is assumed to be 1024
which is common in this workload. So the impact sounds not a big deal.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:36 +00:00
|
|
|
out:
|
2017-05-03 21:52:07 +00:00
|
|
|
for_each_evictable_lru(lru) {
|
|
|
|
int file = is_file_lru(lru);
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
unsigned long lruvec_size;
|
2021-08-20 02:04:21 +00:00
|
|
|
unsigned long low, min;
|
2017-05-03 21:52:07 +00:00
|
|
|
unsigned long scan;
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
|
|
|
|
lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
|
2021-08-20 02:04:21 +00:00
|
|
|
mem_cgroup_protection(sc->target_mem_cgroup, memcg,
|
|
|
|
&min, &low);
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
|
2021-08-20 02:04:21 +00:00
|
|
|
if (min || low) {
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
/*
|
|
|
|
* Scale a cgroup's reclaim pressure by proportioning
|
|
|
|
* its current usage to its memory.low or memory.min
|
|
|
|
* setting.
|
|
|
|
*
|
|
|
|
* This is important, as otherwise scanning aggression
|
|
|
|
* becomes extremely binary -- from nothing as we
|
|
|
|
* approach the memory protection threshold, to totally
|
|
|
|
* nominal as we exceed it. This results in requiring
|
|
|
|
* setting extremely liberal protection thresholds. It
|
|
|
|
* also means we simply get no protection at all if we
|
|
|
|
* set it too low, which is not ideal.
|
mm, memcg: make scan aggression always exclude protection
This patch is an incremental improvement on the existing
memory.{low,min} relative reclaim work to base its scan pressure
calculations on how much protection is available compared to the current
usage, rather than how much the current usage is over some protection
threshold.
This change doesn't change the experience for the user in the normal
case too much. One benefit is that it replaces the (somewhat arbitrary)
100% cutoff with an indefinite slope, which makes it easier to ballpark
a memory.low value.
As well as this, the old methodology doesn't quite apply generically to
machines with varying amounts of physical memory. Let's say we have a
top level cgroup, workload.slice, and another top level cgroup,
system-management.slice. We want to roughly give 12G to
system-management.slice, so on a 32GB machine we set memory.low to 20GB
in workload.slice, and on a 64GB machine we set memory.low to 52GB.
However, because these are relative amounts to the total machine size,
while the amount of memory we want to generally be willing to yield to
system.slice is absolute (12G), we end up putting more pressure on
system.slice just because we have a larger machine and a larger workload
to fill it, which seems fairly unintuitive. With this new behaviour, we
don't end up with this unintended side effect.
Previously the way that memory.low protection works is that if you are
50% over a certain baseline, you get 50% of your normal scan pressure.
This is certainly better than the previous cliff-edge behaviour, but it
can be improved even further by always considering memory under the
currently enforced protection threshold to be out of bounds. This means
that we can set relatively low memory.low thresholds for variable or
bursty workloads while still getting a reasonable level of protection,
whereas with the previous version we may still trivially hit the 100%
clamp. The previous 100% clamp is also somewhat arbitrary, whereas this
one is more concretely based on the currently enforced protection
threshold, which is likely easier to reason about.
There is also a subtle issue with the way that proportional reclaim
worked previously -- it promotes having no memory.low, since it makes
pressure higher during low reclaim. This happens because we base our
scan pressure modulation on how far memory.current is between memory.min
and memory.low, but if memory.low is unset, we only use the overage
method. In most cromulent configurations, this then means that we end
up with *more* pressure than with no memory.low at all when we're in low
reclaim, which is not really very usable or expected.
With this patch, memory.low and memory.min affect reclaim pressure in a
more understandable and composable way. For example, from a user
standpoint, "protected" memory now remains untouchable from a reclaim
aggression standpoint, and users can also have more confidence that
bursty workloads will still receive some amount of guaranteed
protection.
Link: http://lkml.kernel.org/r/20190322160307.GA3316@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:38 +00:00
|
|
|
*
|
|
|
|
* If there is any protection in place, we reduce scan
|
|
|
|
* pressure by how much of the total memory used is
|
|
|
|
* within protection thresholds.
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
*
|
mm, memcg: make memory.emin the baseline for utilisation determination
Roman points out that when when we do the low reclaim pass, we scale the
reclaim pressure relative to position between 0 and the maximum
protection threshold.
However, if the maximum protection is based on memory.elow, and
memory.emin is above zero, this means we still may get binary behaviour
on second-pass low reclaim. This is because we scale starting at 0, not
starting at memory.emin, and since we don't scan at all below emin, we
end up with cliff behaviour.
This should be a fairly uncommon case since usually we don't go into the
second pass, but it makes sense to scale our low reclaim pressure
starting at emin.
You can test this by catting two large sparse files, one in a cgroup
with emin set to some moderate size compared to physical RAM, and
another cgroup without any emin. In both cgroups, set an elow larger
than 50% of physical RAM. The one with emin will have less page
scanning, as reclaim pressure is lower.
Rebase on top of and apply the same idea as what was applied to handle
cgroup_memory=disable properly for the original proportional patch
http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name ("mm,
memcg: Handle cgroup_disable=memory when getting memcg protection").
Link: http://lkml.kernel.org/r/20190201051810.GA18895@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Suggested-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:35 +00:00
|
|
|
* There is one special case: in the first reclaim pass,
|
|
|
|
* we skip over all groups that are within their low
|
|
|
|
* protection. If that fails to reclaim enough pages to
|
|
|
|
* satisfy the reclaim goal, we come back and override
|
|
|
|
* the best-effort low protection. However, we still
|
|
|
|
* ideally want to honor how well-behaved groups are in
|
|
|
|
* that case instead of simply punishing them all
|
|
|
|
* equally. As such, we reclaim them based on how much
|
mm, memcg: make scan aggression always exclude protection
This patch is an incremental improvement on the existing
memory.{low,min} relative reclaim work to base its scan pressure
calculations on how much protection is available compared to the current
usage, rather than how much the current usage is over some protection
threshold.
This change doesn't change the experience for the user in the normal
case too much. One benefit is that it replaces the (somewhat arbitrary)
100% cutoff with an indefinite slope, which makes it easier to ballpark
a memory.low value.
As well as this, the old methodology doesn't quite apply generically to
machines with varying amounts of physical memory. Let's say we have a
top level cgroup, workload.slice, and another top level cgroup,
system-management.slice. We want to roughly give 12G to
system-management.slice, so on a 32GB machine we set memory.low to 20GB
in workload.slice, and on a 64GB machine we set memory.low to 52GB.
However, because these are relative amounts to the total machine size,
while the amount of memory we want to generally be willing to yield to
system.slice is absolute (12G), we end up putting more pressure on
system.slice just because we have a larger machine and a larger workload
to fill it, which seems fairly unintuitive. With this new behaviour, we
don't end up with this unintended side effect.
Previously the way that memory.low protection works is that if you are
50% over a certain baseline, you get 50% of your normal scan pressure.
This is certainly better than the previous cliff-edge behaviour, but it
can be improved even further by always considering memory under the
currently enforced protection threshold to be out of bounds. This means
that we can set relatively low memory.low thresholds for variable or
bursty workloads while still getting a reasonable level of protection,
whereas with the previous version we may still trivially hit the 100%
clamp. The previous 100% clamp is also somewhat arbitrary, whereas this
one is more concretely based on the currently enforced protection
threshold, which is likely easier to reason about.
There is also a subtle issue with the way that proportional reclaim
worked previously -- it promotes having no memory.low, since it makes
pressure higher during low reclaim. This happens because we base our
scan pressure modulation on how far memory.current is between memory.min
and memory.low, but if memory.low is unset, we only use the overage
method. In most cromulent configurations, this then means that we end
up with *more* pressure than with no memory.low at all when we're in low
reclaim, which is not really very usable or expected.
With this patch, memory.low and memory.min affect reclaim pressure in a
more understandable and composable way. For example, from a user
standpoint, "protected" memory now remains untouchable from a reclaim
aggression standpoint, and users can also have more confidence that
bursty workloads will still receive some amount of guaranteed
protection.
Link: http://lkml.kernel.org/r/20190322160307.GA3316@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:38 +00:00
|
|
|
* memory they are using, reducing the scan pressure
|
|
|
|
* again by how much of the total memory used is under
|
|
|
|
* hard protection.
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
*/
|
mm, memcg: make scan aggression always exclude protection
This patch is an incremental improvement on the existing
memory.{low,min} relative reclaim work to base its scan pressure
calculations on how much protection is available compared to the current
usage, rather than how much the current usage is over some protection
threshold.
This change doesn't change the experience for the user in the normal
case too much. One benefit is that it replaces the (somewhat arbitrary)
100% cutoff with an indefinite slope, which makes it easier to ballpark
a memory.low value.
As well as this, the old methodology doesn't quite apply generically to
machines with varying amounts of physical memory. Let's say we have a
top level cgroup, workload.slice, and another top level cgroup,
system-management.slice. We want to roughly give 12G to
system-management.slice, so on a 32GB machine we set memory.low to 20GB
in workload.slice, and on a 64GB machine we set memory.low to 52GB.
However, because these are relative amounts to the total machine size,
while the amount of memory we want to generally be willing to yield to
system.slice is absolute (12G), we end up putting more pressure on
system.slice just because we have a larger machine and a larger workload
to fill it, which seems fairly unintuitive. With this new behaviour, we
don't end up with this unintended side effect.
Previously the way that memory.low protection works is that if you are
50% over a certain baseline, you get 50% of your normal scan pressure.
This is certainly better than the previous cliff-edge behaviour, but it
can be improved even further by always considering memory under the
currently enforced protection threshold to be out of bounds. This means
that we can set relatively low memory.low thresholds for variable or
bursty workloads while still getting a reasonable level of protection,
whereas with the previous version we may still trivially hit the 100%
clamp. The previous 100% clamp is also somewhat arbitrary, whereas this
one is more concretely based on the currently enforced protection
threshold, which is likely easier to reason about.
There is also a subtle issue with the way that proportional reclaim
worked previously -- it promotes having no memory.low, since it makes
pressure higher during low reclaim. This happens because we base our
scan pressure modulation on how far memory.current is between memory.min
and memory.low, but if memory.low is unset, we only use the overage
method. In most cromulent configurations, this then means that we end
up with *more* pressure than with no memory.low at all when we're in low
reclaim, which is not really very usable or expected.
With this patch, memory.low and memory.min affect reclaim pressure in a
more understandable and composable way. For example, from a user
standpoint, "protected" memory now remains untouchable from a reclaim
aggression standpoint, and users can also have more confidence that
bursty workloads will still receive some amount of guaranteed
protection.
Link: http://lkml.kernel.org/r/20190322160307.GA3316@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:38 +00:00
|
|
|
unsigned long cgroup_size = mem_cgroup_size(memcg);
|
2021-08-20 02:04:21 +00:00
|
|
|
unsigned long protection;
|
|
|
|
|
|
|
|
/* memory.low scaling, make sure we retry before OOM */
|
|
|
|
if (!sc->memcg_low_reclaim && low > min) {
|
|
|
|
protection = low;
|
|
|
|
sc->memcg_low_skipped = 1;
|
|
|
|
} else {
|
|
|
|
protection = min;
|
|
|
|
}
|
mm, memcg: make scan aggression always exclude protection
This patch is an incremental improvement on the existing
memory.{low,min} relative reclaim work to base its scan pressure
calculations on how much protection is available compared to the current
usage, rather than how much the current usage is over some protection
threshold.
This change doesn't change the experience for the user in the normal
case too much. One benefit is that it replaces the (somewhat arbitrary)
100% cutoff with an indefinite slope, which makes it easier to ballpark
a memory.low value.
As well as this, the old methodology doesn't quite apply generically to
machines with varying amounts of physical memory. Let's say we have a
top level cgroup, workload.slice, and another top level cgroup,
system-management.slice. We want to roughly give 12G to
system-management.slice, so on a 32GB machine we set memory.low to 20GB
in workload.slice, and on a 64GB machine we set memory.low to 52GB.
However, because these are relative amounts to the total machine size,
while the amount of memory we want to generally be willing to yield to
system.slice is absolute (12G), we end up putting more pressure on
system.slice just because we have a larger machine and a larger workload
to fill it, which seems fairly unintuitive. With this new behaviour, we
don't end up with this unintended side effect.
Previously the way that memory.low protection works is that if you are
50% over a certain baseline, you get 50% of your normal scan pressure.
This is certainly better than the previous cliff-edge behaviour, but it
can be improved even further by always considering memory under the
currently enforced protection threshold to be out of bounds. This means
that we can set relatively low memory.low thresholds for variable or
bursty workloads while still getting a reasonable level of protection,
whereas with the previous version we may still trivially hit the 100%
clamp. The previous 100% clamp is also somewhat arbitrary, whereas this
one is more concretely based on the currently enforced protection
threshold, which is likely easier to reason about.
There is also a subtle issue with the way that proportional reclaim
worked previously -- it promotes having no memory.low, since it makes
pressure higher during low reclaim. This happens because we base our
scan pressure modulation on how far memory.current is between memory.min
and memory.low, but if memory.low is unset, we only use the overage
method. In most cromulent configurations, this then means that we end
up with *more* pressure than with no memory.low at all when we're in low
reclaim, which is not really very usable or expected.
With this patch, memory.low and memory.min affect reclaim pressure in a
more understandable and composable way. For example, from a user
standpoint, "protected" memory now remains untouchable from a reclaim
aggression standpoint, and users can also have more confidence that
bursty workloads will still receive some amount of guaranteed
protection.
Link: http://lkml.kernel.org/r/20190322160307.GA3316@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:38 +00:00
|
|
|
|
|
|
|
/* Avoid TOCTOU with earlier protection check */
|
|
|
|
cgroup_size = max(cgroup_size, protection);
|
|
|
|
|
|
|
|
scan = lruvec_size - lruvec_size * protection /
|
2021-09-09 01:10:08 +00:00
|
|
|
(cgroup_size + 1);
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
|
|
|
|
/*
|
mm, memcg: make scan aggression always exclude protection
This patch is an incremental improvement on the existing
memory.{low,min} relative reclaim work to base its scan pressure
calculations on how much protection is available compared to the current
usage, rather than how much the current usage is over some protection
threshold.
This change doesn't change the experience for the user in the normal
case too much. One benefit is that it replaces the (somewhat arbitrary)
100% cutoff with an indefinite slope, which makes it easier to ballpark
a memory.low value.
As well as this, the old methodology doesn't quite apply generically to
machines with varying amounts of physical memory. Let's say we have a
top level cgroup, workload.slice, and another top level cgroup,
system-management.slice. We want to roughly give 12G to
system-management.slice, so on a 32GB machine we set memory.low to 20GB
in workload.slice, and on a 64GB machine we set memory.low to 52GB.
However, because these are relative amounts to the total machine size,
while the amount of memory we want to generally be willing to yield to
system.slice is absolute (12G), we end up putting more pressure on
system.slice just because we have a larger machine and a larger workload
to fill it, which seems fairly unintuitive. With this new behaviour, we
don't end up with this unintended side effect.
Previously the way that memory.low protection works is that if you are
50% over a certain baseline, you get 50% of your normal scan pressure.
This is certainly better than the previous cliff-edge behaviour, but it
can be improved even further by always considering memory under the
currently enforced protection threshold to be out of bounds. This means
that we can set relatively low memory.low thresholds for variable or
bursty workloads while still getting a reasonable level of protection,
whereas with the previous version we may still trivially hit the 100%
clamp. The previous 100% clamp is also somewhat arbitrary, whereas this
one is more concretely based on the currently enforced protection
threshold, which is likely easier to reason about.
There is also a subtle issue with the way that proportional reclaim
worked previously -- it promotes having no memory.low, since it makes
pressure higher during low reclaim. This happens because we base our
scan pressure modulation on how far memory.current is between memory.min
and memory.low, but if memory.low is unset, we only use the overage
method. In most cromulent configurations, this then means that we end
up with *more* pressure than with no memory.low at all when we're in low
reclaim, which is not really very usable or expected.
With this patch, memory.low and memory.min affect reclaim pressure in a
more understandable and composable way. For example, from a user
standpoint, "protected" memory now remains untouchable from a reclaim
aggression standpoint, and users can also have more confidence that
bursty workloads will still receive some amount of guaranteed
protection.
Link: http://lkml.kernel.org/r/20190322160307.GA3316@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:38 +00:00
|
|
|
* Minimally target SWAP_CLUSTER_MAX pages to keep
|
2020-06-04 23:49:10 +00:00
|
|
|
* reclaim moving forwards, avoiding decrementing
|
mm, memcg: make memory.emin the baseline for utilisation determination
Roman points out that when when we do the low reclaim pass, we scale the
reclaim pressure relative to position between 0 and the maximum
protection threshold.
However, if the maximum protection is based on memory.elow, and
memory.emin is above zero, this means we still may get binary behaviour
on second-pass low reclaim. This is because we scale starting at 0, not
starting at memory.emin, and since we don't scan at all below emin, we
end up with cliff behaviour.
This should be a fairly uncommon case since usually we don't go into the
second pass, but it makes sense to scale our low reclaim pressure
starting at emin.
You can test this by catting two large sparse files, one in a cgroup
with emin set to some moderate size compared to physical RAM, and
another cgroup without any emin. In both cgroups, set an elow larger
than 50% of physical RAM. The one with emin will have less page
scanning, as reclaim pressure is lower.
Rebase on top of and apply the same idea as what was applied to handle
cgroup_memory=disable properly for the original proportional patch
http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name ("mm,
memcg: Handle cgroup_disable=memory when getting memcg protection").
Link: http://lkml.kernel.org/r/20190201051810.GA18895@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Suggested-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:35 +00:00
|
|
|
* sc->priority further than desirable.
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
*/
|
mm, memcg: make scan aggression always exclude protection
This patch is an incremental improvement on the existing
memory.{low,min} relative reclaim work to base its scan pressure
calculations on how much protection is available compared to the current
usage, rather than how much the current usage is over some protection
threshold.
This change doesn't change the experience for the user in the normal
case too much. One benefit is that it replaces the (somewhat arbitrary)
100% cutoff with an indefinite slope, which makes it easier to ballpark
a memory.low value.
As well as this, the old methodology doesn't quite apply generically to
machines with varying amounts of physical memory. Let's say we have a
top level cgroup, workload.slice, and another top level cgroup,
system-management.slice. We want to roughly give 12G to
system-management.slice, so on a 32GB machine we set memory.low to 20GB
in workload.slice, and on a 64GB machine we set memory.low to 52GB.
However, because these are relative amounts to the total machine size,
while the amount of memory we want to generally be willing to yield to
system.slice is absolute (12G), we end up putting more pressure on
system.slice just because we have a larger machine and a larger workload
to fill it, which seems fairly unintuitive. With this new behaviour, we
don't end up with this unintended side effect.
Previously the way that memory.low protection works is that if you are
50% over a certain baseline, you get 50% of your normal scan pressure.
This is certainly better than the previous cliff-edge behaviour, but it
can be improved even further by always considering memory under the
currently enforced protection threshold to be out of bounds. This means
that we can set relatively low memory.low thresholds for variable or
bursty workloads while still getting a reasonable level of protection,
whereas with the previous version we may still trivially hit the 100%
clamp. The previous 100% clamp is also somewhat arbitrary, whereas this
one is more concretely based on the currently enforced protection
threshold, which is likely easier to reason about.
There is also a subtle issue with the way that proportional reclaim
worked previously -- it promotes having no memory.low, since it makes
pressure higher during low reclaim. This happens because we base our
scan pressure modulation on how far memory.current is between memory.min
and memory.low, but if memory.low is unset, we only use the overage
method. In most cromulent configurations, this then means that we end
up with *more* pressure than with no memory.low at all when we're in low
reclaim, which is not really very usable or expected.
With this patch, memory.low and memory.min affect reclaim pressure in a
more understandable and composable way. For example, from a user
standpoint, "protected" memory now remains untouchable from a reclaim
aggression standpoint, and users can also have more confidence that
bursty workloads will still receive some amount of guaranteed
protection.
Link: http://lkml.kernel.org/r/20190322160307.GA3316@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Reviewed-by: Roman Gushchin <guro@fb.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:38 +00:00
|
|
|
scan = max(scan, SWAP_CLUSTER_MAX);
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
} else {
|
|
|
|
scan = lruvec_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
scan >>= sc->priority;
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
|
2017-05-03 21:52:07 +00:00
|
|
|
/*
|
|
|
|
* If the cgroup's already been deleted, make sure to
|
|
|
|
* scrape out the remaining cache.
|
|
|
|
*/
|
|
|
|
if (!scan && !mem_cgroup_online(memcg))
|
mm, memcg: proportional memory.{low,min} reclaim
cgroup v2 introduces two memory protection thresholds: memory.low
(best-effort) and memory.min (hard protection). While they generally do
what they say on the tin, there is a limitation in their implementation
that makes them difficult to use effectively: that cliff behaviour often
manifests when they become eligible for reclaim. This patch implements
more intuitive and usable behaviour, where we gradually mount more
reclaim pressure as cgroups further and further exceed their protection
thresholds.
This cliff edge behaviour happens because we only choose whether or not
to reclaim based on whether the memcg is within its protection limits
(see the use of mem_cgroup_protected in shrink_node), but we don't vary
our reclaim behaviour based on this information. Imagine the following
timeline, with the numbers the lruvec size in this zone:
1. memory.low=1000000, memory.current=999999. 0 pages may be scanned.
2. memory.low=1000000, memory.current=1000000. 0 pages may be scanned.
3. memory.low=1000000, memory.current=1000001. 1000001* pages may be
scanned. (?!)
* Of course, we won't usually scan all available pages in the zone even
without this patch because of scan control priority, over-reclaim
protection, etc. However, as shown by the tests at the end, these
techniques don't sufficiently throttle such an extreme change in input,
so cliff-like behaviour isn't really averted by their existence alone.
Here's an example of how this plays out in practice. At Facebook, we are
trying to protect various workloads from "system" software, like
configuration management tools, metric collectors, etc (see this[0] case
study). In order to find a suitable memory.low value, we start by
determining the expected memory range within which the workload will be
comfortable operating. This isn't an exact science -- memory usage deemed
"comfortable" will vary over time due to user behaviour, differences in
composition of work, etc, etc. As such we need to ballpark memory.low,
but doing this is currently problematic:
1. If we end up setting it too low for the workload, it won't have
*any* effect (see discussion above). The group will receive the full
weight of reclaim and won't have any priority while competing with the
less important system software, as if we had no memory.low configured
at all.
2. Because of this behaviour, we end up erring on the side of setting
it too high, such that the comfort range is reliably covered. However,
protected memory is completely unavailable to the rest of the system,
so we might cause undue memory and IO pressure there when we *know* we
have some elasticity in the workload.
3. Even if we get the value totally right, smack in the middle of the
comfort zone, we get extreme jumps between no pressure and full
pressure that cause unpredictable pressure spikes in the workload due
to the current binary reclaim behaviour.
With this patch, we can set it to our ballpark estimation without too much
worry. Any undesirable behaviour, such as too much or too little reclaim
pressure on the workload or system will be proportional to how far our
estimation is off. This means we can set memory.low much more
conservatively and thus waste less resources *without* the risk of the
workload falling off a cliff if we overshoot.
As a more abstract technical description, this unintuitive behaviour
results in having to give high-priority workloads a large protection
buffer on top of their expected usage to function reliably, as otherwise
we have abrupt periods of dramatically increased memory pressure which
hamper performance. Having to set these thresholds so high wastes
resources and generally works against the principle of work conservation.
In addition, having proportional memory reclaim behaviour has other
benefits. Most notably, before this patch it's basically mandatory to set
memory.low to a higher than desirable value because otherwise as soon as
you exceed memory.low, all protection is lost, and all pages are eligible
to scan again. By contrast, having a gradual ramp in reclaim pressure
means that you now still get some protection when thresholds are exceeded,
which means that one can now be more comfortable setting memory.low to
lower values without worrying that all protection will be lost. This is
important because workingset size is really hard to know exactly,
especially with variable workloads, so at least getting *some* protection
if your workingset size grows larger than you expect increases user
confidence in setting memory.low without a huge buffer on top being
needed.
Thanks a lot to Johannes Weiner and Tejun Heo for their advice and
assistance in thinking about how to make this work better.
In testing these changes, I intended to verify that:
1. Changes in page scanning become gradual and proportional instead of
binary.
To test this, I experimented stepping further and further down
memory.low protection on a workload that floats around 19G workingset
when under memory.low protection, watching page scan rates for the
workload cgroup:
+------------+-----------------+--------------------+--------------+
| memory.low | test (pgscan/s) | control (pgscan/s) | % of control |
+------------+-----------------+--------------------+--------------+
| 21G | 0 | 0 | N/A |
| 17G | 867 | 3799 | 23% |
| 12G | 1203 | 3543 | 34% |
| 8G | 2534 | 3979 | 64% |
| 4G | 3980 | 4147 | 96% |
| 0 | 3799 | 3980 | 95% |
+------------+-----------------+--------------------+--------------+
As you can see, the test kernel (with a kernel containing this
patch) ramps up page scanning significantly more gradually than the
control kernel (without this patch).
2. More gradual ramp up in reclaim aggression doesn't result in
premature OOMs.
To test this, I wrote a script that slowly increments the number of
pages held by stress(1)'s --vm-keep mode until a production system
entered severe overall memory contention. This script runs in a highly
protected slice taking up the majority of available system memory.
Watching vmstat revealed that page scanning continued essentially
nominally between test and control, without causing forward reclaim
progress to become arrested.
[0]: https://facebookmicrosites.github.io/cgroup2/docs/overview.html#case-study-the-fbtax2-project
[akpm@linux-foundation.org: reflow block comments to fit in 80 cols]
[chris@chrisdown.name: handle cgroup_disable=memory when getting memcg protection]
Link: http://lkml.kernel.org/r/20190201045711.GA18302@chrisdown.name
Link: http://lkml.kernel.org/r/20190124014455.GA6396@chrisdown.name
Signed-off-by: Chris Down <chris@chrisdown.name>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 00:58:32 +00:00
|
|
|
scan = min(lruvec_size, SWAP_CLUSTER_MAX);
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
|
2017-05-03 21:52:07 +00:00
|
|
|
switch (scan_balance) {
|
|
|
|
case SCAN_EQUAL:
|
|
|
|
/* Scan lists relative to size */
|
|
|
|
break;
|
|
|
|
case SCAN_FRACT:
|
2013-02-23 00:32:17 +00:00
|
|
|
/*
|
2017-05-03 21:52:07 +00:00
|
|
|
* Scan types proportional to swappiness and
|
|
|
|
* their relative recent reclaim efficiency.
|
2020-02-21 04:04:24 +00:00
|
|
|
* Make sure we don't miss the last page on
|
|
|
|
* the offlined memory cgroups because of a
|
|
|
|
* round-off error.
|
2013-02-23 00:32:17 +00:00
|
|
|
*/
|
2020-02-21 04:04:24 +00:00
|
|
|
scan = mem_cgroup_online(memcg) ?
|
|
|
|
div64_u64(scan * fraction[file], denominator) :
|
|
|
|
DIV64_U64_ROUND_UP(scan * fraction[file],
|
2018-10-26 22:03:27 +00:00
|
|
|
denominator);
|
2017-05-03 21:52:07 +00:00
|
|
|
break;
|
|
|
|
case SCAN_FILE:
|
|
|
|
case SCAN_ANON:
|
|
|
|
/* Scan one type exclusively */
|
2020-04-02 04:10:15 +00:00
|
|
|
if ((scan_balance == SCAN_FILE) != file)
|
2017-05-03 21:52:07 +00:00
|
|
|
scan = 0;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
/* Look ma, no brain */
|
|
|
|
BUG();
|
2013-02-23 00:32:17 +00:00
|
|
|
}
|
2017-05-03 21:52:07 +00:00
|
|
|
|
|
|
|
nr[lru] = scan;
|
vmscan: prevent get_scan_ratio() rounding errors
get_scan_ratio() calculates percentage and if the percentage is < 1%, it
will round percentage down to 0% and cause we completely ignore scanning
anon/file pages to reclaim memory even the total anon/file pages are very
big.
To avoid underflow, we don't use percentage, instead we directly calculate
how many pages should be scaned. In this way, we should get several
scanned pages for < 1% percent.
This has some benefits:
1. increase our calculation precision
2. making our scan more smoothly. Without this, if percent[x] is
underflow, shrink_zone() doesn't scan any pages and suddenly it scans
all pages when priority is zero. With this, even priority isn't zero,
shrink_zone() gets chance to scan some pages.
Note, this patch doesn't really change logics, but just increase
precision. For system with a lot of memory, this might slightly changes
behavior. For example, in a sequential file read workload, without the
patch, we don't swap any anon pages. With it, if anon memory size is
bigger than 16G, we will see one anon page swapped. The 16G is calculated
as PAGE_SIZE * priority(4096) * (fp/ap). fp/ap is assumed to be 1024
which is common in this workload. So the impact sounds not a big deal.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-24 21:32:36 +00:00
|
|
|
}
|
2009-06-16 22:32:29 +00:00
|
|
|
}
|
2008-10-19 03:26:32 +00:00
|
|
|
|
2021-09-02 21:59:23 +00:00
|
|
|
/*
|
|
|
|
* Anonymous LRU management is a waste if there is
|
|
|
|
* ultimately no way to reclaim the memory.
|
|
|
|
*/
|
|
|
|
static bool can_age_anon_pages(struct pglist_data *pgdat,
|
|
|
|
struct scan_control *sc)
|
|
|
|
{
|
|
|
|
/* Aging the anon LRU is valuable if swap is present: */
|
|
|
|
if (total_swap_pages > 0)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
/* Also valuable if anon pages can be demoted: */
|
|
|
|
return can_demote(pgdat->node_id, sc);
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
#ifdef CONFIG_LRU_GEN
|
|
|
|
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
#ifdef CONFIG_LRU_GEN_ENABLED
|
|
|
|
DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
|
|
|
|
#define get_cap(cap) static_branch_likely(&lru_gen_caps[cap])
|
|
|
|
#else
|
|
|
|
DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
|
|
|
|
#define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap])
|
|
|
|
#endif
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* shorthand helpers
|
|
|
|
******************************************************************************/
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
#define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset))
|
|
|
|
|
|
|
|
#define DEFINE_MAX_SEQ(lruvec) \
|
|
|
|
unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
|
|
|
|
|
|
|
|
#define DEFINE_MIN_SEQ(lruvec) \
|
|
|
|
unsigned long min_seq[ANON_AND_FILE] = { \
|
|
|
|
READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \
|
|
|
|
READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
#define for_each_gen_type_zone(gen, type, zone) \
|
|
|
|
for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \
|
|
|
|
for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \
|
|
|
|
for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
#define get_memcg_gen(seq) ((seq) % MEMCG_NR_GENS)
|
|
|
|
#define get_memcg_bin(bin) ((bin) % MEMCG_NR_BINS)
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
{
|
|
|
|
struct pglist_data *pgdat = NODE_DATA(nid);
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
if (memcg) {
|
|
|
|
struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
|
|
|
|
|
2022-11-16 01:38:08 +00:00
|
|
|
/* see the comment in mem_cgroup_lruvec() */
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
if (!lruvec->pgdat)
|
|
|
|
lruvec->pgdat = pgdat;
|
|
|
|
|
|
|
|
return lruvec;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
VM_WARN_ON_ONCE(!mem_cgroup_disabled());
|
|
|
|
|
2022-11-16 01:38:08 +00:00
|
|
|
return &pgdat->__lruvec;
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(lruvec);
|
|
|
|
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
if (!sc->may_swap)
|
|
|
|
return 0;
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
if (!can_demote(pgdat->node_id, sc) &&
|
|
|
|
mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
return mem_cgroup_swappiness(memcg);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int get_nr_gens(struct lruvec *lruvec, int type)
|
|
|
|
{
|
|
|
|
return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
|
|
|
|
{
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
/* see the comment on lru_gen_folio */
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
|
|
|
|
get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
|
|
|
|
get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
|
|
|
|
}
|
|
|
|
|
2023-01-18 00:18:23 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* Bloom filters
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
|
|
|
|
* n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
|
|
|
|
* bits in a bitmap, k is the number of hash functions and n is the number of
|
|
|
|
* inserted items.
|
|
|
|
*
|
|
|
|
* Page table walkers use one of the two filters to reduce their search space.
|
|
|
|
* To get rid of non-leaf entries that no longer have enough leaf entries, the
|
|
|
|
* aging uses the double-buffering technique to flip to the other filter each
|
|
|
|
* time it produces a new generation. For non-leaf entries that have enough
|
|
|
|
* leaf entries, the aging carries them over to the next generation in
|
|
|
|
* walk_pmd_range(); the eviction also report them when walking the rmap
|
|
|
|
* in lru_gen_look_around().
|
|
|
|
*
|
|
|
|
* For future optimizations:
|
|
|
|
* 1. It's not necessary to keep both filters all the time. The spare one can be
|
|
|
|
* freed after the RCU grace period and reallocated if needed again.
|
|
|
|
* 2. And when reallocating, it's worth scaling its size according to the number
|
|
|
|
* of inserted entries in the other filter, to reduce the memory overhead on
|
|
|
|
* small systems and false positives on large systems.
|
|
|
|
* 3. Jenkins' hash function is an alternative to Knuth's.
|
|
|
|
*/
|
|
|
|
#define BLOOM_FILTER_SHIFT 15
|
|
|
|
|
|
|
|
static inline int filter_gen_from_seq(unsigned long seq)
|
|
|
|
{
|
|
|
|
return seq % NR_BLOOM_FILTERS;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void get_item_key(void *item, int *key)
|
|
|
|
{
|
|
|
|
u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
|
|
|
|
|
|
|
|
BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
|
|
|
|
|
|
|
|
key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
|
|
|
|
key[1] = hash >> BLOOM_FILTER_SHIFT;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
|
|
|
|
{
|
|
|
|
int key[2];
|
|
|
|
unsigned long *filter;
|
|
|
|
int gen = filter_gen_from_seq(seq);
|
|
|
|
|
|
|
|
filter = READ_ONCE(lruvec->mm_state.filters[gen]);
|
|
|
|
if (!filter)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
get_item_key(item, key);
|
|
|
|
|
|
|
|
return test_bit(key[0], filter) && test_bit(key[1], filter);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void update_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
|
|
|
|
{
|
|
|
|
int key[2];
|
|
|
|
unsigned long *filter;
|
|
|
|
int gen = filter_gen_from_seq(seq);
|
|
|
|
|
|
|
|
filter = READ_ONCE(lruvec->mm_state.filters[gen]);
|
|
|
|
if (!filter)
|
|
|
|
return;
|
|
|
|
|
|
|
|
get_item_key(item, key);
|
|
|
|
|
|
|
|
if (!test_bit(key[0], filter))
|
|
|
|
set_bit(key[0], filter);
|
|
|
|
if (!test_bit(key[1], filter))
|
|
|
|
set_bit(key[1], filter);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq)
|
|
|
|
{
|
|
|
|
unsigned long *filter;
|
|
|
|
int gen = filter_gen_from_seq(seq);
|
|
|
|
|
|
|
|
filter = lruvec->mm_state.filters[gen];
|
|
|
|
if (filter) {
|
|
|
|
bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
|
|
|
|
__GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
|
|
|
|
WRITE_ONCE(lruvec->mm_state.filters[gen], filter);
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* mm_struct list
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
static struct lru_gen_mm_list mm_list = {
|
|
|
|
.fifo = LIST_HEAD_INIT(mm_list.fifo),
|
|
|
|
.lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
|
|
|
|
};
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
if (memcg)
|
|
|
|
return &memcg->mm_list;
|
|
|
|
#endif
|
|
|
|
VM_WARN_ON_ONCE(!mem_cgroup_disabled());
|
|
|
|
|
|
|
|
return &mm_list;
|
|
|
|
}
|
|
|
|
|
|
|
|
void lru_gen_add_mm(struct mm_struct *mm)
|
|
|
|
{
|
|
|
|
int nid;
|
|
|
|
struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
|
|
|
|
struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
VM_WARN_ON_ONCE(mm->lru_gen.memcg);
|
|
|
|
mm->lru_gen.memcg = memcg;
|
|
|
|
#endif
|
|
|
|
spin_lock(&mm_list->lock);
|
|
|
|
|
|
|
|
for_each_node_state(nid, N_MEMORY) {
|
|
|
|
struct lruvec *lruvec = get_lruvec(memcg, nid);
|
|
|
|
|
|
|
|
/* the first addition since the last iteration */
|
|
|
|
if (lruvec->mm_state.tail == &mm_list->fifo)
|
|
|
|
lruvec->mm_state.tail = &mm->lru_gen.list;
|
|
|
|
}
|
|
|
|
|
|
|
|
list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
|
|
|
|
|
|
|
|
spin_unlock(&mm_list->lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
void lru_gen_del_mm(struct mm_struct *mm)
|
|
|
|
{
|
|
|
|
int nid;
|
|
|
|
struct lru_gen_mm_list *mm_list;
|
|
|
|
struct mem_cgroup *memcg = NULL;
|
|
|
|
|
|
|
|
if (list_empty(&mm->lru_gen.list))
|
|
|
|
return;
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
memcg = mm->lru_gen.memcg;
|
|
|
|
#endif
|
|
|
|
mm_list = get_mm_list(memcg);
|
|
|
|
|
|
|
|
spin_lock(&mm_list->lock);
|
|
|
|
|
|
|
|
for_each_node(nid) {
|
|
|
|
struct lruvec *lruvec = get_lruvec(memcg, nid);
|
|
|
|
|
|
|
|
/* where the last iteration ended (exclusive) */
|
|
|
|
if (lruvec->mm_state.tail == &mm->lru_gen.list)
|
|
|
|
lruvec->mm_state.tail = lruvec->mm_state.tail->next;
|
|
|
|
|
|
|
|
/* where the current iteration continues (inclusive) */
|
|
|
|
if (lruvec->mm_state.head != &mm->lru_gen.list)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
lruvec->mm_state.head = lruvec->mm_state.head->next;
|
|
|
|
/* the deletion ends the current iteration */
|
|
|
|
if (lruvec->mm_state.head == &mm_list->fifo)
|
|
|
|
WRITE_ONCE(lruvec->mm_state.seq, lruvec->mm_state.seq + 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
list_del_init(&mm->lru_gen.list);
|
|
|
|
|
|
|
|
spin_unlock(&mm_list->lock);
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
mem_cgroup_put(mm->lru_gen.memcg);
|
|
|
|
mm->lru_gen.memcg = NULL;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
void lru_gen_migrate_mm(struct mm_struct *mm)
|
|
|
|
{
|
|
|
|
struct mem_cgroup *memcg;
|
|
|
|
struct task_struct *task = rcu_dereference_protected(mm->owner, true);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(task->mm != mm);
|
|
|
|
lockdep_assert_held(&task->alloc_lock);
|
|
|
|
|
|
|
|
/* for mm_update_next_owner() */
|
|
|
|
if (mem_cgroup_disabled())
|
|
|
|
return;
|
|
|
|
|
2023-01-16 03:44:05 +00:00
|
|
|
/* migration can happen before addition */
|
|
|
|
if (!mm->lru_gen.memcg)
|
|
|
|
return;
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
rcu_read_lock();
|
|
|
|
memcg = mem_cgroup_from_task(task);
|
|
|
|
rcu_read_unlock();
|
|
|
|
if (memcg == mm->lru_gen.memcg)
|
|
|
|
return;
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
|
|
|
|
|
|
|
|
lru_gen_del_mm(mm);
|
|
|
|
lru_gen_add_mm(mm);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
int hist;
|
|
|
|
|
|
|
|
lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
|
|
|
|
|
|
|
|
if (walk) {
|
|
|
|
hist = lru_hist_from_seq(walk->max_seq);
|
|
|
|
|
|
|
|
for (i = 0; i < NR_MM_STATS; i++) {
|
|
|
|
WRITE_ONCE(lruvec->mm_state.stats[hist][i],
|
|
|
|
lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]);
|
|
|
|
walk->mm_stats[i] = 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (NR_HIST_GENS > 1 && last) {
|
|
|
|
hist = lru_hist_from_seq(lruvec->mm_state.seq + 1);
|
|
|
|
|
|
|
|
for (i = 0; i < NR_MM_STATS; i++)
|
|
|
|
WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
|
|
|
|
{
|
|
|
|
int type;
|
|
|
|
unsigned long size = 0;
|
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
|
|
|
|
int key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
|
|
|
|
|
|
|
|
if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
clear_bit(key, &mm->lru_gen.bitmap);
|
|
|
|
|
|
|
|
for (type = !walk->can_swap; type < ANON_AND_FILE; type++) {
|
|
|
|
size += type ? get_mm_counter(mm, MM_FILEPAGES) :
|
|
|
|
get_mm_counter(mm, MM_ANONPAGES) +
|
|
|
|
get_mm_counter(mm, MM_SHMEMPAGES);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (size < MIN_LRU_BATCH)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
return !mmget_not_zero(mm);
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
|
|
|
|
struct mm_struct **iter)
|
|
|
|
{
|
|
|
|
bool first = false;
|
|
|
|
bool last = true;
|
|
|
|
struct mm_struct *mm = NULL;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
|
|
|
|
struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* There are four interesting cases for this page table walker:
|
|
|
|
* 1. It tries to start a new iteration of mm_list with a stale max_seq;
|
|
|
|
* there is nothing left to do.
|
|
|
|
* 2. It's the first of the current generation, and it needs to reset
|
|
|
|
* the Bloom filter for the next generation.
|
|
|
|
* 3. It reaches the end of mm_list, and it needs to increment
|
|
|
|
* mm_state->seq; the iteration is done.
|
|
|
|
* 4. It's the last of the current generation, and it needs to reset the
|
|
|
|
* mm stats counters for the next generation.
|
|
|
|
*/
|
|
|
|
spin_lock(&mm_list->lock);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq);
|
|
|
|
VM_WARN_ON_ONCE(*iter && mm_state->seq > walk->max_seq);
|
|
|
|
VM_WARN_ON_ONCE(*iter && !mm_state->nr_walkers);
|
|
|
|
|
|
|
|
if (walk->max_seq <= mm_state->seq) {
|
|
|
|
if (!*iter)
|
|
|
|
last = false;
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!mm_state->nr_walkers) {
|
|
|
|
VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
|
|
|
|
|
|
|
|
mm_state->head = mm_list->fifo.next;
|
|
|
|
first = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
while (!mm && mm_state->head != &mm_list->fifo) {
|
|
|
|
mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
|
|
|
|
|
|
|
|
mm_state->head = mm_state->head->next;
|
|
|
|
|
|
|
|
/* force scan for those added after the last iteration */
|
|
|
|
if (!mm_state->tail || mm_state->tail == &mm->lru_gen.list) {
|
|
|
|
mm_state->tail = mm_state->head;
|
|
|
|
walk->force_scan = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (should_skip_mm(mm, walk))
|
|
|
|
mm = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (mm_state->head == &mm_list->fifo)
|
|
|
|
WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
|
|
|
|
done:
|
|
|
|
if (*iter && !mm)
|
|
|
|
mm_state->nr_walkers--;
|
|
|
|
if (!*iter && mm)
|
|
|
|
mm_state->nr_walkers++;
|
|
|
|
|
|
|
|
if (mm_state->nr_walkers)
|
|
|
|
last = false;
|
|
|
|
|
|
|
|
if (*iter || last)
|
|
|
|
reset_mm_stats(lruvec, walk, last);
|
|
|
|
|
|
|
|
spin_unlock(&mm_list->lock);
|
|
|
|
|
|
|
|
if (mm && first)
|
|
|
|
reset_bloom_filter(lruvec, walk->max_seq + 1);
|
|
|
|
|
|
|
|
if (*iter)
|
|
|
|
mmput_async(*iter);
|
|
|
|
|
|
|
|
*iter = mm;
|
|
|
|
|
|
|
|
return last;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
|
|
|
|
{
|
|
|
|
bool success = false;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
|
|
|
|
struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
|
|
|
|
|
|
|
|
spin_lock(&mm_list->lock);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq);
|
|
|
|
|
|
|
|
if (max_seq > mm_state->seq && !mm_state->nr_walkers) {
|
|
|
|
VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
|
|
|
|
|
|
|
|
WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
|
|
|
|
reset_mm_stats(lruvec, NULL, true);
|
|
|
|
success = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_unlock(&mm_list->lock);
|
|
|
|
|
|
|
|
return success;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* refault feedback loop
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* A feedback loop based on Proportional-Integral-Derivative (PID) controller.
|
|
|
|
*
|
|
|
|
* The P term is refaulted/(evicted+protected) from a tier in the generation
|
|
|
|
* currently being evicted; the I term is the exponential moving average of the
|
|
|
|
* P term over the generations previously evicted, using the smoothing factor
|
|
|
|
* 1/2; the D term isn't supported.
|
|
|
|
*
|
|
|
|
* The setpoint (SP) is always the first tier of one type; the process variable
|
|
|
|
* (PV) is either any tier of the other type or any other tier of the same
|
|
|
|
* type.
|
|
|
|
*
|
|
|
|
* The error is the difference between the SP and the PV; the correction is to
|
|
|
|
* turn off protection when SP>PV or turn on protection when SP<PV.
|
|
|
|
*
|
|
|
|
* For future optimizations:
|
|
|
|
* 1. The D term may discount the other two terms over time so that long-lived
|
|
|
|
* generations can resist stale information.
|
|
|
|
*/
|
|
|
|
struct ctrl_pos {
|
|
|
|
unsigned long refaulted;
|
|
|
|
unsigned long total;
|
|
|
|
int gain;
|
|
|
|
};
|
|
|
|
|
|
|
|
static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
|
|
|
|
struct ctrl_pos *pos)
|
|
|
|
{
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
int hist = lru_hist_from_seq(lrugen->min_seq[type]);
|
|
|
|
|
|
|
|
pos->refaulted = lrugen->avg_refaulted[type][tier] +
|
|
|
|
atomic_long_read(&lrugen->refaulted[hist][type][tier]);
|
|
|
|
pos->total = lrugen->avg_total[type][tier] +
|
|
|
|
atomic_long_read(&lrugen->evicted[hist][type][tier]);
|
|
|
|
if (tier)
|
|
|
|
pos->total += lrugen->protected[hist][type][tier - 1];
|
|
|
|
pos->gain = gain;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
|
|
|
|
{
|
|
|
|
int hist, tier;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
|
|
|
|
unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
|
|
|
|
|
|
|
|
lockdep_assert_held(&lruvec->lru_lock);
|
|
|
|
|
|
|
|
if (!carryover && !clear)
|
|
|
|
return;
|
|
|
|
|
|
|
|
hist = lru_hist_from_seq(seq);
|
|
|
|
|
|
|
|
for (tier = 0; tier < MAX_NR_TIERS; tier++) {
|
|
|
|
if (carryover) {
|
|
|
|
unsigned long sum;
|
|
|
|
|
|
|
|
sum = lrugen->avg_refaulted[type][tier] +
|
|
|
|
atomic_long_read(&lrugen->refaulted[hist][type][tier]);
|
|
|
|
WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
|
|
|
|
|
|
|
|
sum = lrugen->avg_total[type][tier] +
|
|
|
|
atomic_long_read(&lrugen->evicted[hist][type][tier]);
|
|
|
|
if (tier)
|
|
|
|
sum += lrugen->protected[hist][type][tier - 1];
|
|
|
|
WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (clear) {
|
|
|
|
atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
|
|
|
|
atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
|
|
|
|
if (tier)
|
|
|
|
WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
* Return true if the PV has a limited number of refaults or a lower
|
|
|
|
* refaulted/total than the SP.
|
|
|
|
*/
|
|
|
|
return pv->refaulted < MIN_LRU_BATCH ||
|
|
|
|
pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
|
|
|
|
(sp->refaulted + 1) * pv->total * pv->gain;
|
|
|
|
}
|
|
|
|
|
|
|
|
/******************************************************************************
|
|
|
|
* the aging
|
|
|
|
******************************************************************************/
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
/* promote pages accessed through page tables */
|
|
|
|
static int folio_update_gen(struct folio *folio, int gen)
|
|
|
|
{
|
|
|
|
unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
|
|
|
|
VM_WARN_ON_ONCE(!rcu_read_lock_held());
|
|
|
|
|
|
|
|
do {
|
|
|
|
/* lru_gen_del_folio() has isolated this page? */
|
|
|
|
if (!(old_flags & LRU_GEN_MASK)) {
|
2022-09-02 19:45:57 +00:00
|
|
|
/* for shrink_folio_list() */
|
2022-09-18 08:00:04 +00:00
|
|
|
new_flags = old_flags | BIT(PG_referenced);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
|
|
|
|
new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
|
|
|
|
} while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
|
|
|
|
|
|
|
|
return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
/* protect pages accessed multiple times through file descriptors */
|
|
|
|
static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
|
|
|
|
{
|
|
|
|
int type = folio_is_file_lru(folio);
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
|
|
|
|
unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
|
|
|
|
|
|
|
|
do {
|
2022-09-18 08:00:04 +00:00
|
|
|
new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
|
|
|
|
/* folio_update_gen() has promoted this page? */
|
|
|
|
if (new_gen >= 0 && new_gen != old_gen)
|
|
|
|
return new_gen;
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
new_gen = (old_gen + 1) % MAX_NR_GENS;
|
|
|
|
|
|
|
|
new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
|
|
|
|
new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
|
|
|
|
/* for folio_end_writeback() */
|
|
|
|
if (reclaiming)
|
|
|
|
new_flags |= BIT(PG_reclaim);
|
|
|
|
} while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
|
|
|
|
|
|
|
|
lru_gen_update_size(lruvec, folio, old_gen, new_gen);
|
|
|
|
|
|
|
|
return new_gen;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
|
|
|
|
int old_gen, int new_gen)
|
|
|
|
{
|
|
|
|
int type = folio_is_file_lru(folio);
|
|
|
|
int zone = folio_zonenum(folio);
|
|
|
|
int delta = folio_nr_pages(folio);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
|
|
|
|
VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
|
|
|
|
|
|
|
|
walk->batched++;
|
|
|
|
|
|
|
|
walk->nr_pages[old_gen][type][zone] -= delta;
|
|
|
|
walk->nr_pages[new_gen][type][zone] += delta;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
|
|
|
|
{
|
|
|
|
int gen, type, zone;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
walk->batched = 0;
|
|
|
|
|
|
|
|
for_each_gen_type_zone(gen, type, zone) {
|
|
|
|
enum lru_list lru = type * LRU_INACTIVE_FILE;
|
|
|
|
int delta = walk->nr_pages[gen][type][zone];
|
|
|
|
|
|
|
|
if (!delta)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
walk->nr_pages[gen][type][zone] = 0;
|
|
|
|
WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
|
|
|
|
lrugen->nr_pages[gen][type][zone] + delta);
|
|
|
|
|
|
|
|
if (lru_gen_is_active(lruvec, gen))
|
|
|
|
lru += LRU_ACTIVE;
|
|
|
|
__update_lru_size(lruvec, lru, zone, delta);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
|
|
|
|
{
|
|
|
|
struct address_space *mapping;
|
|
|
|
struct vm_area_struct *vma = args->vma;
|
|
|
|
struct lru_gen_mm_walk *walk = args->private;
|
|
|
|
|
|
|
|
if (!vma_is_accessible(vma))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
if (is_vm_hugetlb_page(vma))
|
|
|
|
return true;
|
|
|
|
|
2022-12-30 21:52:51 +00:00
|
|
|
if (!vma_has_recency(vma))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL))
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
return true;
|
|
|
|
|
|
|
|
if (vma == get_gate_vma(vma->vm_mm))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
if (vma_is_anonymous(vma))
|
|
|
|
return !walk->can_swap;
|
|
|
|
|
|
|
|
if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
mapping = vma->vm_file->f_mapping;
|
|
|
|
if (mapping_unevictable(mapping))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
if (shmem_mapping(mapping))
|
|
|
|
return !walk->can_swap;
|
|
|
|
|
|
|
|
/* to exclude special mappings like dax, etc. */
|
|
|
|
return !mapping->a_ops->read_folio;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Some userspace memory allocators map many single-page VMAs. Instead of
|
|
|
|
* returning back to the PGD table for each of such VMAs, finish an entire PMD
|
|
|
|
* table to reduce zigzags and improve cache performance.
|
|
|
|
*/
|
|
|
|
static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
|
|
|
|
unsigned long *vm_start, unsigned long *vm_end)
|
|
|
|
{
|
|
|
|
unsigned long start = round_up(*vm_end, size);
|
|
|
|
unsigned long end = (start | ~mask) + 1;
|
2022-09-06 19:49:05 +00:00
|
|
|
VMA_ITERATOR(vmi, args->mm, start);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(mask & size);
|
|
|
|
VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
|
|
|
|
|
2022-09-06 19:49:05 +00:00
|
|
|
for_each_vma(vmi, args->vma) {
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
if (end && end <= args->vma->vm_start)
|
|
|
|
return false;
|
|
|
|
|
2022-09-06 19:49:05 +00:00
|
|
|
if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
*vm_start = max(start, args->vma->vm_start);
|
|
|
|
*vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
|
|
|
|
{
|
|
|
|
unsigned long pfn = pte_pfn(pte);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
|
|
|
|
|
|
|
|
if (!pte_present(pte) || is_zero_pfn(pfn))
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
if (WARN_ON_ONCE(!pfn_valid(pfn)))
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
return pfn;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
|
|
|
|
static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
|
|
|
|
{
|
|
|
|
unsigned long pfn = pmd_pfn(pmd);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
|
|
|
|
|
|
|
|
if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
if (WARN_ON_ONCE(pmd_devmap(pmd)))
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
if (WARN_ON_ONCE(!pfn_valid(pfn)))
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
return pfn;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
struct pglist_data *pgdat, bool can_swap)
|
2022-09-18 08:00:04 +00:00
|
|
|
{
|
|
|
|
struct folio *folio;
|
|
|
|
|
|
|
|
/* try to avoid unnecessary memory loads */
|
|
|
|
if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
folio = pfn_folio(pfn);
|
|
|
|
if (folio_nid(folio) != pgdat->node_id)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
if (folio_memcg_rcu(folio) != memcg)
|
|
|
|
return NULL;
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
/* file VMAs can contain anon pages from COW */
|
|
|
|
if (!folio_is_file_lru(folio) && !can_swap)
|
|
|
|
return NULL;
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
return folio;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
static bool suitable_to_scan(int total, int young)
|
|
|
|
{
|
|
|
|
int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
|
|
|
|
|
|
|
|
/* suitable if the average number of young PTEs per cacheline is >=1 */
|
|
|
|
return young * n >= total;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
|
|
|
|
struct mm_walk *args)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
pte_t *pte;
|
|
|
|
spinlock_t *ptl;
|
|
|
|
unsigned long addr;
|
|
|
|
int total = 0;
|
|
|
|
int young = 0;
|
|
|
|
struct lru_gen_mm_walk *walk = args->private;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
|
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
|
|
|
|
int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(pmd_leaf(*pmd));
|
|
|
|
|
|
|
|
ptl = pte_lockptr(args->mm, pmd);
|
|
|
|
if (!spin_trylock(ptl))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
arch_enter_lazy_mmu_mode();
|
|
|
|
|
|
|
|
pte = pte_offset_map(pmd, start & PMD_MASK);
|
|
|
|
restart:
|
|
|
|
for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
|
|
|
|
unsigned long pfn;
|
|
|
|
struct folio *folio;
|
|
|
|
|
|
|
|
total++;
|
|
|
|
walk->mm_stats[MM_LEAF_TOTAL]++;
|
|
|
|
|
|
|
|
pfn = get_pte_pfn(pte[i], args->vma, addr);
|
|
|
|
if (pfn == -1)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (!pte_young(pte[i])) {
|
|
|
|
walk->mm_stats[MM_LEAF_OLD]++;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
|
|
|
|
if (!folio)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
|
|
|
|
VM_WARN_ON_ONCE(true);
|
|
|
|
|
|
|
|
young++;
|
|
|
|
walk->mm_stats[MM_LEAF_YOUNG]++;
|
|
|
|
|
|
|
|
if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
|
|
|
|
!(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
|
|
|
|
!folio_test_swapcache(folio)))
|
|
|
|
folio_mark_dirty(folio);
|
|
|
|
|
|
|
|
old_gen = folio_update_gen(folio, new_gen);
|
|
|
|
if (old_gen >= 0 && old_gen != new_gen)
|
|
|
|
update_batch_size(walk, folio, old_gen, new_gen);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
|
|
|
|
goto restart;
|
|
|
|
|
|
|
|
pte_unmap(pte);
|
|
|
|
|
|
|
|
arch_leave_lazy_mmu_mode();
|
|
|
|
spin_unlock(ptl);
|
|
|
|
|
|
|
|
return suitable_to_scan(total, young);
|
|
|
|
}
|
|
|
|
|
|
|
|
#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
|
2023-01-18 00:18:26 +00:00
|
|
|
static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
|
|
|
|
struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
{
|
|
|
|
int i;
|
|
|
|
pmd_t *pmd;
|
|
|
|
spinlock_t *ptl;
|
|
|
|
struct lru_gen_mm_walk *walk = args->private;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
|
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
|
|
|
|
int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(pud_leaf(*pud));
|
|
|
|
|
|
|
|
/* try to batch at most 1+MIN_LRU_BATCH+1 entries */
|
2023-01-18 00:18:26 +00:00
|
|
|
if (*first == -1) {
|
|
|
|
*first = addr;
|
|
|
|
bitmap_zero(bitmap, MIN_LRU_BATCH);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-01-18 00:18:26 +00:00
|
|
|
i = addr == -1 ? 0 : pmd_index(addr) - pmd_index(*first);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
if (i && i <= MIN_LRU_BATCH) {
|
|
|
|
__set_bit(i - 1, bitmap);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-01-18 00:18:26 +00:00
|
|
|
pmd = pmd_offset(pud, *first);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
ptl = pmd_lockptr(args->mm, pmd);
|
|
|
|
if (!spin_trylock(ptl))
|
|
|
|
goto done;
|
|
|
|
|
|
|
|
arch_enter_lazy_mmu_mode();
|
|
|
|
|
|
|
|
do {
|
|
|
|
unsigned long pfn;
|
|
|
|
struct folio *folio;
|
2023-01-18 00:18:26 +00:00
|
|
|
|
|
|
|
/* don't round down the first address */
|
|
|
|
addr = i ? (*first & PMD_MASK) + i * PMD_SIZE : *first;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
pfn = get_pmd_pfn(pmd[i], vma, addr);
|
|
|
|
if (pfn == -1)
|
|
|
|
goto next;
|
|
|
|
|
|
|
|
if (!pmd_trans_huge(pmd[i])) {
|
2023-01-18 00:18:26 +00:00
|
|
|
if (arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG))
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
pmdp_test_and_clear_young(vma, addr, pmd + i);
|
|
|
|
goto next;
|
|
|
|
}
|
|
|
|
|
|
|
|
folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
|
|
|
|
if (!folio)
|
|
|
|
goto next;
|
|
|
|
|
|
|
|
if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
|
|
|
|
goto next;
|
|
|
|
|
|
|
|
walk->mm_stats[MM_LEAF_YOUNG]++;
|
|
|
|
|
|
|
|
if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
|
|
|
|
!(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
|
|
|
|
!folio_test_swapcache(folio)))
|
|
|
|
folio_mark_dirty(folio);
|
|
|
|
|
|
|
|
old_gen = folio_update_gen(folio, new_gen);
|
|
|
|
if (old_gen >= 0 && old_gen != new_gen)
|
|
|
|
update_batch_size(walk, folio, old_gen, new_gen);
|
|
|
|
next:
|
|
|
|
i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
|
|
|
|
} while (i <= MIN_LRU_BATCH);
|
|
|
|
|
|
|
|
arch_leave_lazy_mmu_mode();
|
|
|
|
spin_unlock(ptl);
|
|
|
|
done:
|
2023-01-18 00:18:26 +00:00
|
|
|
*first = -1;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
}
|
|
|
|
#else
|
2023-01-18 00:18:26 +00:00
|
|
|
static void walk_pmd_range_locked(pud_t *pud, unsigned long addr, struct vm_area_struct *vma,
|
|
|
|
struct mm_walk *args, unsigned long *bitmap, unsigned long *first)
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
{
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
|
|
|
|
struct mm_walk *args)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
pmd_t *pmd;
|
|
|
|
unsigned long next;
|
|
|
|
unsigned long addr;
|
|
|
|
struct vm_area_struct *vma;
|
2023-01-18 00:18:26 +00:00
|
|
|
unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)];
|
|
|
|
unsigned long first = -1;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
struct lru_gen_mm_walk *walk = args->private;
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(pud_leaf(*pud));
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Finish an entire PMD in two passes: the first only reaches to PTE
|
|
|
|
* tables to avoid taking the PMD lock; the second, if necessary, takes
|
|
|
|
* the PMD lock to clear the accessed bit in PMD entries.
|
|
|
|
*/
|
|
|
|
pmd = pmd_offset(pud, start & PUD_MASK);
|
|
|
|
restart:
|
|
|
|
/* walk_pte_range() may call get_next_vma() */
|
|
|
|
vma = args->vma;
|
|
|
|
for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
|
2020-11-26 16:20:28 +00:00
|
|
|
pmd_t val = pmdp_get_lockless(pmd + i);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
next = pmd_addr_end(addr, end);
|
|
|
|
|
|
|
|
if (!pmd_present(val) || is_huge_zero_pmd(val)) {
|
|
|
|
walk->mm_stats[MM_LEAF_TOTAL]++;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
|
|
|
|
if (pmd_trans_huge(val)) {
|
|
|
|
unsigned long pfn = pmd_pfn(val);
|
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
|
|
|
|
|
|
|
|
walk->mm_stats[MM_LEAF_TOTAL]++;
|
|
|
|
|
|
|
|
if (!pmd_young(val)) {
|
|
|
|
walk->mm_stats[MM_LEAF_OLD]++;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* try to avoid unnecessary memory loads */
|
|
|
|
if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
|
|
|
|
continue;
|
|
|
|
|
2023-01-18 00:18:26 +00:00
|
|
|
walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
walk->mm_stats[MM_NONLEAF_TOTAL]++;
|
|
|
|
|
2023-01-18 00:18:26 +00:00
|
|
|
if (arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG)) {
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
if (!pmd_young(val))
|
|
|
|
continue;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
2023-01-18 00:18:26 +00:00
|
|
|
walk_pmd_range_locked(pud, addr, vma, args, bitmap, &first);
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
}
|
2022-11-23 06:45:10 +00:00
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
if (!walk->force_scan && !test_bloom_filter(walk->lruvec, walk->max_seq, pmd + i))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
walk->mm_stats[MM_NONLEAF_FOUND]++;
|
|
|
|
|
|
|
|
if (!walk_pte_range(&val, addr, next, args))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
walk->mm_stats[MM_NONLEAF_ADDED]++;
|
|
|
|
|
|
|
|
/* carry over to the next generation */
|
|
|
|
update_bloom_filter(walk->lruvec, walk->max_seq + 1, pmd + i);
|
|
|
|
}
|
|
|
|
|
2023-01-18 00:18:26 +00:00
|
|
|
walk_pmd_range_locked(pud, -1, vma, args, bitmap, &first);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
|
|
|
|
struct mm_walk *args)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
pud_t *pud;
|
|
|
|
unsigned long addr;
|
|
|
|
unsigned long next;
|
|
|
|
struct lru_gen_mm_walk *walk = args->private;
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(p4d_leaf(*p4d));
|
|
|
|
|
|
|
|
pud = pud_offset(p4d, start & P4D_MASK);
|
|
|
|
restart:
|
|
|
|
for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
|
|
|
|
pud_t val = READ_ONCE(pud[i]);
|
|
|
|
|
|
|
|
next = pud_addr_end(addr, end);
|
|
|
|
|
|
|
|
if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
walk_pmd_range(&val, addr, next, args);
|
|
|
|
|
|
|
|
/* a racy check to curtail the waiting time */
|
|
|
|
if (wq_has_sleeper(&walk->lruvec->mm_state.wait))
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
|
|
|
|
end = (addr | ~PUD_MASK) + 1;
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
|
|
|
|
goto restart;
|
|
|
|
|
|
|
|
end = round_up(end, P4D_SIZE);
|
|
|
|
done:
|
|
|
|
if (!end || !args->vma)
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
walk->next_addr = max(end, args->vma->vm_start);
|
|
|
|
|
|
|
|
return -EAGAIN;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
|
|
|
|
{
|
|
|
|
static const struct mm_walk_ops mm_walk_ops = {
|
|
|
|
.test_walk = should_skip_vma,
|
|
|
|
.p4d_entry = walk_pud_range,
|
|
|
|
};
|
|
|
|
|
|
|
|
int err;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
|
|
|
|
walk->next_addr = FIRST_USER_ADDRESS;
|
|
|
|
|
|
|
|
do {
|
|
|
|
err = -EBUSY;
|
|
|
|
|
|
|
|
/* folio_update_gen() requires stable folio_memcg() */
|
|
|
|
if (!mem_cgroup_trylock_pages(memcg))
|
|
|
|
break;
|
|
|
|
|
|
|
|
/* the caller might be holding the lock for write */
|
|
|
|
if (mmap_read_trylock(mm)) {
|
|
|
|
err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
|
|
|
|
|
|
|
|
mmap_read_unlock(mm);
|
|
|
|
}
|
|
|
|
|
|
|
|
mem_cgroup_unlock_pages();
|
|
|
|
|
|
|
|
if (walk->batched) {
|
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
|
|
|
reset_batch_size(lruvec, walk);
|
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
cond_resched();
|
|
|
|
} while (err == -EAGAIN);
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat, bool force_alloc)
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
{
|
|
|
|
struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
|
|
|
|
|
|
|
|
if (pgdat && current_is_kswapd()) {
|
|
|
|
VM_WARN_ON_ONCE(walk);
|
|
|
|
|
|
|
|
walk = &pgdat->mm_walk;
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
} else if (!walk && force_alloc) {
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
VM_WARN_ON_ONCE(current_is_kswapd());
|
|
|
|
|
|
|
|
walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
|
|
|
|
}
|
|
|
|
|
|
|
|
current->reclaim_state->mm_walk = walk;
|
|
|
|
|
|
|
|
return walk;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void clear_mm_walk(void)
|
|
|
|
{
|
|
|
|
struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
|
|
|
|
VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
|
|
|
|
|
|
|
|
current->reclaim_state->mm_walk = NULL;
|
|
|
|
|
|
|
|
if (!current_is_kswapd())
|
|
|
|
kfree(walk);
|
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
{
|
2022-09-18 08:00:09 +00:00
|
|
|
int zone;
|
|
|
|
int remaining = MAX_LRU_BATCH;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
2022-09-18 08:00:09 +00:00
|
|
|
int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
|
|
|
|
|
|
|
|
if (type == LRU_GEN_ANON && !can_swap)
|
|
|
|
goto done;
|
|
|
|
|
|
|
|
/* prevent cold/hot inversion if force_scan is true */
|
|
|
|
for (zone = 0; zone < MAX_NR_ZONES; zone++) {
|
2022-12-22 04:19:00 +00:00
|
|
|
struct list_head *head = &lrugen->folios[old_gen][type][zone];
|
2022-09-18 08:00:09 +00:00
|
|
|
|
|
|
|
while (!list_empty(head)) {
|
|
|
|
struct folio *folio = lru_to_folio(head);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
new_gen = folio_inc_gen(lruvec, folio, false);
|
2022-12-22 04:19:00 +00:00
|
|
|
list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]);
|
2022-09-18 08:00:09 +00:00
|
|
|
|
|
|
|
if (!--remaining)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
done:
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
reset_ctrl_pos(lruvec, type, true);
|
|
|
|
WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
|
2022-09-18 08:00:09 +00:00
|
|
|
|
|
|
|
return true;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
|
|
|
|
{
|
|
|
|
int gen, type, zone;
|
|
|
|
bool success = false;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
DEFINE_MIN_SEQ(lruvec);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
|
|
|
|
|
|
|
|
/* find the oldest populated generation */
|
|
|
|
for (type = !can_swap; type < ANON_AND_FILE; type++) {
|
|
|
|
while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
|
|
|
|
gen = lru_gen_from_seq(min_seq[type]);
|
|
|
|
|
|
|
|
for (zone = 0; zone < MAX_NR_ZONES; zone++) {
|
2022-12-22 04:19:00 +00:00
|
|
|
if (!list_empty(&lrugen->folios[gen][type][zone]))
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
goto next;
|
|
|
|
}
|
|
|
|
|
|
|
|
min_seq[type]++;
|
|
|
|
}
|
|
|
|
next:
|
|
|
|
;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
/* see the comment on lru_gen_folio */
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
if (can_swap) {
|
|
|
|
min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
|
|
|
|
min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (type = !can_swap; type < ANON_AND_FILE; type++) {
|
|
|
|
if (min_seq[type] == lrugen->min_seq[type])
|
|
|
|
continue;
|
|
|
|
|
|
|
|
reset_ctrl_pos(lruvec, type, true);
|
|
|
|
WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
|
|
|
|
success = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return success;
|
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan)
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
{
|
|
|
|
int prev, next;
|
|
|
|
int type, zone;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
|
|
|
|
|
|
|
|
for (type = ANON_AND_FILE - 1; type >= 0; type--) {
|
|
|
|
if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
|
|
|
|
continue;
|
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
while (!inc_min_seq(lruvec, type, can_swap)) {
|
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
|
|
|
cond_resched();
|
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
|
|
|
}
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Update the active/inactive LRU sizes for compatibility. Both sides of
|
|
|
|
* the current max_seq need to be covered, since max_seq+1 can overlap
|
|
|
|
* with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
|
|
|
|
* overlap, cold/hot inversion happens.
|
|
|
|
*/
|
|
|
|
prev = lru_gen_from_seq(lrugen->max_seq - 1);
|
|
|
|
next = lru_gen_from_seq(lrugen->max_seq + 1);
|
|
|
|
|
|
|
|
for (type = 0; type < ANON_AND_FILE; type++) {
|
|
|
|
for (zone = 0; zone < MAX_NR_ZONES; zone++) {
|
|
|
|
enum lru_list lru = type * LRU_INACTIVE_FILE;
|
|
|
|
long delta = lrugen->nr_pages[prev][type][zone] -
|
|
|
|
lrugen->nr_pages[next][type][zone];
|
|
|
|
|
|
|
|
if (!delta)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
__update_lru_size(lruvec, lru, zone, delta);
|
|
|
|
__update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
for (type = 0; type < ANON_AND_FILE; type++)
|
|
|
|
reset_ctrl_pos(lruvec, type, false);
|
|
|
|
|
2022-09-18 08:00:08 +00:00
|
|
|
WRITE_ONCE(lrugen->timestamps[next], jiffies);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
/* make sure preceding modifications appear */
|
|
|
|
smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
|
2022-09-18 08:00:09 +00:00
|
|
|
struct scan_control *sc, bool can_swap, bool force_scan)
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
{
|
|
|
|
bool success;
|
|
|
|
struct lru_gen_mm_walk *walk;
|
|
|
|
struct mm_struct *mm = NULL;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
|
|
|
|
|
|
|
|
/* see the comment in iterate_mm_list() */
|
|
|
|
if (max_seq <= READ_ONCE(lruvec->mm_state.seq)) {
|
|
|
|
success = false;
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the hardware doesn't automatically set the accessed bit, fallback
|
|
|
|
* to lru_gen_look_around(), which only clears the accessed bit in a
|
|
|
|
* handful of PTEs. Spreading the work out over a period of time usually
|
|
|
|
* is less efficient, but it avoids bursty page faults.
|
|
|
|
*/
|
2022-12-22 04:19:06 +00:00
|
|
|
if (!arch_has_hw_pte_young() || !get_cap(LRU_GEN_MM_WALK)) {
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
success = iterate_mm_list_nowalk(lruvec, max_seq);
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
walk = set_mm_walk(NULL, true);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
if (!walk) {
|
|
|
|
success = iterate_mm_list_nowalk(lruvec, max_seq);
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
|
|
|
walk->lruvec = lruvec;
|
|
|
|
walk->max_seq = max_seq;
|
|
|
|
walk->can_swap = can_swap;
|
2022-09-18 08:00:09 +00:00
|
|
|
walk->force_scan = force_scan;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
do {
|
|
|
|
success = iterate_mm_list(lruvec, walk, &mm);
|
|
|
|
if (mm)
|
|
|
|
walk_mm(lruvec, mm, walk);
|
|
|
|
|
|
|
|
cond_resched();
|
|
|
|
} while (mm);
|
|
|
|
done:
|
|
|
|
if (!success) {
|
|
|
|
if (sc->priority <= DEF_PRIORITY - 2)
|
|
|
|
wait_event_killable(lruvec->mm_state.wait,
|
|
|
|
max_seq < READ_ONCE(lrugen->max_seq));
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
return false;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(max_seq != READ_ONCE(lrugen->max_seq));
|
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
inc_max_seq(lruvec, can_swap, force_scan);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
/* either this sees any waiters or they will see updated max_seq */
|
|
|
|
if (wq_has_sleeper(&lruvec->mm_state.wait))
|
|
|
|
wake_up_all(&lruvec->mm_state.wait);
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2023-01-18 00:18:21 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* working set protection
|
|
|
|
******************************************************************************/
|
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
static bool lruvec_is_sizable(struct lruvec *lruvec, struct scan_control *sc)
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
{
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
int gen, type, zone;
|
|
|
|
unsigned long total = 0;
|
|
|
|
bool can_swap = get_swappiness(lruvec, sc);
|
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
DEFINE_MAX_SEQ(lruvec);
|
|
|
|
DEFINE_MIN_SEQ(lruvec);
|
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
for (type = !can_swap; type < ANON_AND_FILE; type++) {
|
|
|
|
unsigned long seq;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
for (seq = min_seq[type]; seq <= max_seq; seq++) {
|
|
|
|
gen = lru_gen_from_seq(seq);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
for (zone = 0; zone < MAX_NR_ZONES; zone++)
|
|
|
|
total += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
|
|
|
|
}
|
|
|
|
}
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
/* whether the size is big enough to be helpful */
|
|
|
|
return mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
|
|
|
|
}
|
2022-09-18 08:00:08 +00:00
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
static bool lruvec_is_reclaimable(struct lruvec *lruvec, struct scan_control *sc,
|
|
|
|
unsigned long min_ttl)
|
|
|
|
{
|
|
|
|
int gen;
|
|
|
|
unsigned long birth;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
DEFINE_MIN_SEQ(lruvec);
|
2022-09-18 08:00:08 +00:00
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
/* see the comment on lru_gen_folio */
|
|
|
|
gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
|
|
|
|
birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
|
2022-09-18 08:00:08 +00:00
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
if (time_is_after_jiffies(birth + min_ttl))
|
|
|
|
return false;
|
2022-09-18 08:00:08 +00:00
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
if (!lruvec_is_sizable(lruvec, sc))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
mem_cgroup_calculate_protection(NULL, memcg);
|
|
|
|
|
|
|
|
return !mem_cgroup_below_min(NULL, memcg);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:08 +00:00
|
|
|
/* to protect the working set of the last N jiffies */
|
|
|
|
static unsigned long lru_gen_min_ttl __read_mostly;
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
struct mem_cgroup *memcg;
|
2022-09-18 08:00:08 +00:00
|
|
|
unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(!current_is_kswapd());
|
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
/* check the order to exclude compaction-induced reclaim */
|
|
|
|
if (!min_ttl || sc->order || sc->priority == DEF_PRIORITY)
|
mm: multi-gen LRU: optimize multiple memcgs
When multiple memcgs are available, it is possible to use generations as a
frame of reference to make better choices and improve overall performance
under global memory pressure. This patch adds a basic optimization to
select memcgs that can drop single-use unmapped clean pages first. Doing
so reduces the chance of going into the aging path or swapping, which can
be costly.
A typical example that benefits from this optimization is a server running
mixed types of workloads, e.g., heavy anon workload in one memcg and heavy
buffered I/O workload in the other.
Though this optimization can be applied to both kswapd and direct reclaim,
it is only added to kswapd to keep the patchset manageable. Later
improvements may cover the direct reclaim path.
While ensuring certain fairness to all eligible memcgs, proportional scans
of individual memcgs also require proper backoff to avoid overshooting
their aggregate reclaim target by too much. Otherwise it can cause high
direct reclaim latency. The conditions for backoff are:
1. At low priorities, for direct reclaim, if aging fairness or direct
reclaim latency is at risk, i.e., aging one memcg multiple times or
swapping after the target is met.
2. At high priorities, for global reclaim, if per-zone free pages are
above respective watermarks.
Server benchmark results:
Mixed workloads:
fio (buffered I/O): +[19, 21]%
IOPS BW
patch1-8: 1880k 7343MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[119, 123]%
Ops/sec KB/sec
patch1-8: 862768.65 33514.68
patch1-9: 1911022.12 74234.54
Mixed workloads:
fio (buffered I/O): +[75, 77]%
IOPS BW
5.19-rc1: 1279k 4996MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[13, 15]%
Ops/sec KB/sec
5.19-rc1: 1673524.04 65008.87
patch1-9: 1911022.12 74234.54
Configurations:
(changes since patch 6)
cat mixed.sh
modprobe brd rd_nr=2 rd_size=56623104
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
mkfs.ext4 /dev/ram1
mount -t ext4 /dev/ram1 /mnt
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
fio -name=mglru --numjobs=36 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=90m --group_reporting &
pid=$!
sleep 200
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
kill -INT $pid
wait
Client benchmark results:
no change (CONFIG_MEMCG=n)
Link: https://lkml.kernel.org/r/20220918080010.2920238-10-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:06 +00:00
|
|
|
return;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
memcg = mem_cgroup_iter(NULL, NULL, NULL);
|
|
|
|
do {
|
|
|
|
struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
|
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
if (lruvec_is_reclaimable(lruvec, sc, min_ttl)) {
|
|
|
|
mem_cgroup_iter_break(NULL, memcg);
|
|
|
|
return;
|
|
|
|
}
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
cond_resched();
|
|
|
|
} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
2022-09-18 08:00:08 +00:00
|
|
|
/*
|
|
|
|
* The main goal is to OOM kill if every generation from all memcgs is
|
|
|
|
* younger than min_ttl. However, another possibility is all memcgs are
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
* either too small or below min.
|
2022-09-18 08:00:08 +00:00
|
|
|
*/
|
|
|
|
if (mutex_trylock(&oom_lock)) {
|
|
|
|
struct oom_control oc = {
|
|
|
|
.gfp_mask = sc->gfp_mask,
|
|
|
|
};
|
|
|
|
|
|
|
|
out_of_memory(&oc);
|
|
|
|
|
|
|
|
mutex_unlock(&oom_lock);
|
|
|
|
}
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
}
|
|
|
|
|
2023-01-18 00:18:22 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* rmap/PT walk feedback
|
|
|
|
******************************************************************************/
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
/*
|
2022-09-02 19:45:57 +00:00
|
|
|
* This function exploits spatial locality when shrink_folio_list() walks the
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
* rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
|
|
|
|
* the scan was done cacheline efficiently, it adds the PMD entry pointing to
|
|
|
|
* the PTE table to the Bloom filter. This forms a feedback loop between the
|
|
|
|
* eviction and the aging.
|
2022-09-18 08:00:04 +00:00
|
|
|
*/
|
|
|
|
void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
unsigned long start;
|
|
|
|
unsigned long end;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
struct lru_gen_mm_walk *walk;
|
|
|
|
int young = 0;
|
2023-01-18 00:18:27 +00:00
|
|
|
pte_t *pte = pvmw->pte;
|
|
|
|
unsigned long addr = pvmw->address;
|
2022-09-18 08:00:04 +00:00
|
|
|
struct folio *folio = pfn_folio(pvmw->pfn);
|
|
|
|
struct mem_cgroup *memcg = folio_memcg(folio);
|
|
|
|
struct pglist_data *pgdat = folio_pgdat(folio);
|
|
|
|
struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
|
|
|
|
DEFINE_MAX_SEQ(lruvec);
|
|
|
|
int old_gen, new_gen = lru_gen_from_seq(max_seq);
|
|
|
|
|
|
|
|
lockdep_assert_held(pvmw->ptl);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
|
|
|
|
|
|
|
|
if (spin_is_contended(pvmw->ptl))
|
|
|
|
return;
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
/* avoid taking the LRU lock under the PTL when possible */
|
|
|
|
walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
|
|
|
|
|
2023-01-18 00:18:27 +00:00
|
|
|
start = max(addr & PMD_MASK, pvmw->vma->vm_start);
|
|
|
|
end = min(addr | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
|
2022-09-18 08:00:04 +00:00
|
|
|
|
|
|
|
if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
|
2023-01-18 00:18:27 +00:00
|
|
|
if (addr - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
|
2022-09-18 08:00:04 +00:00
|
|
|
end = start + MIN_LRU_BATCH * PAGE_SIZE;
|
2023-01-18 00:18:27 +00:00
|
|
|
else if (end - addr < MIN_LRU_BATCH * PAGE_SIZE / 2)
|
2022-09-18 08:00:04 +00:00
|
|
|
start = end - MIN_LRU_BATCH * PAGE_SIZE;
|
|
|
|
else {
|
2023-01-18 00:18:27 +00:00
|
|
|
start = addr - MIN_LRU_BATCH * PAGE_SIZE / 2;
|
|
|
|
end = addr + MIN_LRU_BATCH * PAGE_SIZE / 2;
|
2022-09-18 08:00:04 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-01-18 00:18:27 +00:00
|
|
|
/* folio_update_gen() requires stable folio_memcg() */
|
|
|
|
if (!mem_cgroup_trylock_pages(memcg))
|
|
|
|
return;
|
2022-09-18 08:00:04 +00:00
|
|
|
|
|
|
|
arch_enter_lazy_mmu_mode();
|
|
|
|
|
2023-01-18 00:18:27 +00:00
|
|
|
pte -= (addr - start) / PAGE_SIZE;
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
|
|
|
|
unsigned long pfn;
|
|
|
|
|
|
|
|
pfn = get_pte_pfn(pte[i], pvmw->vma, addr);
|
|
|
|
if (pfn == -1)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (!pte_young(pte[i]))
|
|
|
|
continue;
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
folio = get_pfn_folio(pfn, memcg, pgdat, !walk || walk->can_swap);
|
2022-09-18 08:00:04 +00:00
|
|
|
if (!folio)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (!ptep_test_and_clear_young(pvmw->vma, addr, pte + i))
|
|
|
|
VM_WARN_ON_ONCE(true);
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
young++;
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
|
|
|
|
!(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
|
|
|
|
!folio_test_swapcache(folio)))
|
|
|
|
folio_mark_dirty(folio);
|
|
|
|
|
2023-01-18 00:18:27 +00:00
|
|
|
if (walk) {
|
|
|
|
old_gen = folio_update_gen(folio, new_gen);
|
|
|
|
if (old_gen >= 0 && old_gen != new_gen)
|
|
|
|
update_batch_size(walk, folio, old_gen, new_gen);
|
|
|
|
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
old_gen = folio_lru_gen(folio);
|
|
|
|
if (old_gen < 0)
|
|
|
|
folio_set_referenced(folio);
|
|
|
|
else if (old_gen != new_gen)
|
2023-01-18 00:18:27 +00:00
|
|
|
folio_activate(folio);
|
2022-09-18 08:00:04 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
arch_leave_lazy_mmu_mode();
|
2023-01-18 00:18:27 +00:00
|
|
|
mem_cgroup_unlock_pages();
|
2022-09-18 08:00:04 +00:00
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
/* feedback from rmap walkers to page table walkers */
|
|
|
|
if (suitable_to_scan(i, young))
|
|
|
|
update_bloom_filter(lruvec, max_seq, pvmw->pmd);
|
2022-09-18 08:00:04 +00:00
|
|
|
}
|
|
|
|
|
2023-01-18 00:18:24 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* memcg LRU
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
/* see the comment on MEMCG_NR_GENS */
|
|
|
|
enum {
|
|
|
|
MEMCG_LRU_NOP,
|
|
|
|
MEMCG_LRU_HEAD,
|
|
|
|
MEMCG_LRU_TAIL,
|
|
|
|
MEMCG_LRU_OLD,
|
|
|
|
MEMCG_LRU_YOUNG,
|
|
|
|
};
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
|
|
|
|
static int lru_gen_memcg_seg(struct lruvec *lruvec)
|
|
|
|
{
|
|
|
|
return READ_ONCE(lruvec->lrugen.seg);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lru_gen_rotate_memcg(struct lruvec *lruvec, int op)
|
|
|
|
{
|
|
|
|
int seg;
|
|
|
|
int old, new;
|
|
|
|
int bin = get_random_u32_below(MEMCG_NR_BINS);
|
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(lruvec);
|
|
|
|
|
|
|
|
spin_lock(&pgdat->memcg_lru.lock);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
|
|
|
|
|
|
|
|
seg = 0;
|
|
|
|
new = old = lruvec->lrugen.gen;
|
|
|
|
|
|
|
|
/* see the comment on MEMCG_NR_GENS */
|
|
|
|
if (op == MEMCG_LRU_HEAD)
|
|
|
|
seg = MEMCG_LRU_HEAD;
|
|
|
|
else if (op == MEMCG_LRU_TAIL)
|
|
|
|
seg = MEMCG_LRU_TAIL;
|
|
|
|
else if (op == MEMCG_LRU_OLD)
|
|
|
|
new = get_memcg_gen(pgdat->memcg_lru.seq);
|
|
|
|
else if (op == MEMCG_LRU_YOUNG)
|
|
|
|
new = get_memcg_gen(pgdat->memcg_lru.seq + 1);
|
|
|
|
else
|
|
|
|
VM_WARN_ON_ONCE(true);
|
|
|
|
|
|
|
|
hlist_nulls_del_rcu(&lruvec->lrugen.list);
|
|
|
|
|
|
|
|
if (op == MEMCG_LRU_HEAD || op == MEMCG_LRU_OLD)
|
|
|
|
hlist_nulls_add_head_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
|
|
|
|
else
|
|
|
|
hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[new][bin]);
|
|
|
|
|
|
|
|
pgdat->memcg_lru.nr_memcgs[old]--;
|
|
|
|
pgdat->memcg_lru.nr_memcgs[new]++;
|
|
|
|
|
|
|
|
lruvec->lrugen.gen = new;
|
|
|
|
WRITE_ONCE(lruvec->lrugen.seg, seg);
|
|
|
|
|
|
|
|
if (!pgdat->memcg_lru.nr_memcgs[old] && old == get_memcg_gen(pgdat->memcg_lru.seq))
|
|
|
|
WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
|
|
|
|
|
|
|
|
spin_unlock(&pgdat->memcg_lru.lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
void lru_gen_online_memcg(struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
int gen;
|
|
|
|
int nid;
|
|
|
|
int bin = get_random_u32_below(MEMCG_NR_BINS);
|
|
|
|
|
|
|
|
for_each_node(nid) {
|
|
|
|
struct pglist_data *pgdat = NODE_DATA(nid);
|
|
|
|
struct lruvec *lruvec = get_lruvec(memcg, nid);
|
|
|
|
|
|
|
|
spin_lock(&pgdat->memcg_lru.lock);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(!hlist_nulls_unhashed(&lruvec->lrugen.list));
|
|
|
|
|
|
|
|
gen = get_memcg_gen(pgdat->memcg_lru.seq);
|
|
|
|
|
|
|
|
hlist_nulls_add_tail_rcu(&lruvec->lrugen.list, &pgdat->memcg_lru.fifo[gen][bin]);
|
|
|
|
pgdat->memcg_lru.nr_memcgs[gen]++;
|
|
|
|
|
|
|
|
lruvec->lrugen.gen = gen;
|
|
|
|
|
|
|
|
spin_unlock(&pgdat->memcg_lru.lock);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void lru_gen_offline_memcg(struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
int nid;
|
|
|
|
|
|
|
|
for_each_node(nid) {
|
|
|
|
struct lruvec *lruvec = get_lruvec(memcg, nid);
|
|
|
|
|
|
|
|
lru_gen_rotate_memcg(lruvec, MEMCG_LRU_OLD);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void lru_gen_release_memcg(struct mem_cgroup *memcg)
|
|
|
|
{
|
|
|
|
int gen;
|
|
|
|
int nid;
|
|
|
|
|
|
|
|
for_each_node(nid) {
|
|
|
|
struct pglist_data *pgdat = NODE_DATA(nid);
|
|
|
|
struct lruvec *lruvec = get_lruvec(memcg, nid);
|
|
|
|
|
|
|
|
spin_lock(&pgdat->memcg_lru.lock);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(hlist_nulls_unhashed(&lruvec->lrugen.list));
|
|
|
|
|
|
|
|
gen = lruvec->lrugen.gen;
|
|
|
|
|
|
|
|
hlist_nulls_del_rcu(&lruvec->lrugen.list);
|
|
|
|
pgdat->memcg_lru.nr_memcgs[gen]--;
|
|
|
|
|
|
|
|
if (!pgdat->memcg_lru.nr_memcgs[gen] && gen == get_memcg_gen(pgdat->memcg_lru.seq))
|
|
|
|
WRITE_ONCE(pgdat->memcg_lru.seq, pgdat->memcg_lru.seq + 1);
|
|
|
|
|
|
|
|
spin_unlock(&pgdat->memcg_lru.lock);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void lru_gen_soft_reclaim(struct lruvec *lruvec)
|
|
|
|
{
|
|
|
|
/* see the comment on MEMCG_NR_GENS */
|
|
|
|
if (lru_gen_memcg_seg(lruvec) != MEMCG_LRU_HEAD)
|
|
|
|
lru_gen_rotate_memcg(lruvec, MEMCG_LRU_HEAD);
|
|
|
|
}
|
|
|
|
|
|
|
|
#else /* !CONFIG_MEMCG */
|
|
|
|
|
|
|
|
static int lru_gen_memcg_seg(struct lruvec *lruvec)
|
|
|
|
{
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* the eviction
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
static bool sort_folio(struct lruvec *lruvec, struct folio *folio, int tier_idx)
|
|
|
|
{
|
|
|
|
bool success;
|
|
|
|
int gen = folio_lru_gen(folio);
|
|
|
|
int type = folio_is_file_lru(folio);
|
|
|
|
int zone = folio_zonenum(folio);
|
|
|
|
int delta = folio_nr_pages(folio);
|
|
|
|
int refs = folio_lru_refs(folio);
|
|
|
|
int tier = lru_tier_from_refs(refs);
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
|
|
|
|
|
|
|
|
/* unevictable */
|
|
|
|
if (!folio_evictable(folio)) {
|
|
|
|
success = lru_gen_del_folio(lruvec, folio, true);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(!success, folio);
|
|
|
|
folio_set_unevictable(folio);
|
|
|
|
lruvec_add_folio(lruvec, folio);
|
|
|
|
__count_vm_events(UNEVICTABLE_PGCULLED, delta);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* dirty lazyfree */
|
|
|
|
if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
|
|
|
|
success = lru_gen_del_folio(lruvec, folio, true);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(!success, folio);
|
|
|
|
folio_set_swapbacked(folio);
|
|
|
|
lruvec_add_folio_tail(lruvec, folio);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:04 +00:00
|
|
|
/* promoted */
|
|
|
|
if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
|
2022-12-22 04:19:00 +00:00
|
|
|
list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
|
2022-09-18 08:00:04 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
/* protected */
|
|
|
|
if (tier > tier_idx) {
|
|
|
|
int hist = lru_hist_from_seq(lrugen->min_seq[type]);
|
|
|
|
|
|
|
|
gen = folio_inc_gen(lruvec, folio, false);
|
2022-12-22 04:19:00 +00:00
|
|
|
list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
|
|
|
|
lrugen->protected[hist][type][tier - 1] + delta);
|
|
|
|
__mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* waiting for writeback */
|
|
|
|
if (folio_test_locked(folio) || folio_test_writeback(folio) ||
|
|
|
|
(type == LRU_GEN_FILE && folio_test_dirty(folio))) {
|
|
|
|
gen = folio_inc_gen(lruvec, folio, true);
|
2022-12-22 04:19:00 +00:00
|
|
|
list_move(&folio->lru, &lrugen->folios[gen][type][zone]);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
bool success;
|
|
|
|
|
|
|
|
/* swapping inhibited */
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
if (!(sc->gfp_mask & __GFP_IO) &&
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
(folio_test_dirty(folio) ||
|
|
|
|
(folio_test_anon(folio) && !folio_test_swapcache(folio))))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
/* raced with release_pages() */
|
|
|
|
if (!folio_try_get(folio))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
/* raced with another isolation */
|
|
|
|
if (!folio_test_clear_lru(folio)) {
|
|
|
|
folio_put(folio);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* see the comment on MAX_NR_TIERS */
|
|
|
|
if (!folio_test_referenced(folio))
|
|
|
|
set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
/* for shrink_folio_list() */
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
folio_clear_reclaim(folio);
|
|
|
|
folio_clear_referenced(folio);
|
|
|
|
|
|
|
|
success = lru_gen_del_folio(lruvec, folio, true);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(!success, folio);
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
|
|
|
|
int type, int tier, struct list_head *list)
|
|
|
|
{
|
|
|
|
int gen, zone;
|
|
|
|
enum vm_event_item item;
|
|
|
|
int sorted = 0;
|
|
|
|
int scanned = 0;
|
|
|
|
int isolated = 0;
|
|
|
|
int remaining = MAX_LRU_BATCH;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(!list_empty(list));
|
|
|
|
|
|
|
|
if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
gen = lru_gen_from_seq(lrugen->min_seq[type]);
|
|
|
|
|
|
|
|
for (zone = sc->reclaim_idx; zone >= 0; zone--) {
|
|
|
|
LIST_HEAD(moved);
|
|
|
|
int skipped = 0;
|
2022-12-22 04:19:00 +00:00
|
|
|
struct list_head *head = &lrugen->folios[gen][type][zone];
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
while (!list_empty(head)) {
|
|
|
|
struct folio *folio = lru_to_folio(head);
|
|
|
|
int delta = folio_nr_pages(folio);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
|
|
|
|
|
|
|
|
scanned += delta;
|
|
|
|
|
|
|
|
if (sort_folio(lruvec, folio, tier))
|
|
|
|
sorted += delta;
|
|
|
|
else if (isolate_folio(lruvec, folio, sc)) {
|
|
|
|
list_add(&folio->lru, list);
|
|
|
|
isolated += delta;
|
|
|
|
} else {
|
|
|
|
list_move(&folio->lru, &moved);
|
|
|
|
skipped += delta;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (skipped) {
|
|
|
|
list_splice(&moved, head);
|
|
|
|
__count_zid_vm_events(PGSCAN_SKIP, zone, skipped);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!remaining || isolated >= MIN_LRU_BATCH)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2022-10-26 18:01:33 +00:00
|
|
|
item = PGSCAN_KSWAPD + reclaimer_offset();
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
if (!cgroup_reclaim(sc)) {
|
|
|
|
__count_vm_events(item, isolated);
|
|
|
|
__count_vm_events(PGREFILL, sorted);
|
|
|
|
}
|
|
|
|
__count_memcg_events(memcg, item, isolated);
|
|
|
|
__count_memcg_events(memcg, PGREFILL, sorted);
|
|
|
|
__count_vm_events(PGSCAN_ANON + type, isolated);
|
|
|
|
|
|
|
|
/*
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
* There might not be eligible folios due to reclaim_idx. Check the
|
|
|
|
* remaining to prevent livelock if it's not making progress.
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
*/
|
|
|
|
return isolated || !remaining ? scanned : 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int get_tier_idx(struct lruvec *lruvec, int type)
|
|
|
|
{
|
|
|
|
int tier;
|
|
|
|
struct ctrl_pos sp, pv;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* To leave a margin for fluctuations, use a larger gain factor (1:2).
|
|
|
|
* This value is chosen because any other tier would have at least twice
|
|
|
|
* as many refaults as the first tier.
|
|
|
|
*/
|
|
|
|
read_ctrl_pos(lruvec, type, 0, 1, &sp);
|
|
|
|
for (tier = 1; tier < MAX_NR_TIERS; tier++) {
|
|
|
|
read_ctrl_pos(lruvec, type, tier, 2, &pv);
|
|
|
|
if (!positive_ctrl_err(&sp, &pv))
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return tier - 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
|
|
|
|
{
|
|
|
|
int type, tier;
|
|
|
|
struct ctrl_pos sp, pv;
|
|
|
|
int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Compare the first tier of anon with that of file to determine which
|
|
|
|
* type to scan. Also need to compare other tiers of the selected type
|
|
|
|
* with the first tier of the other type to determine the last tier (of
|
|
|
|
* the selected type) to evict.
|
|
|
|
*/
|
|
|
|
read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
|
|
|
|
read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
|
|
|
|
type = positive_ctrl_err(&sp, &pv);
|
|
|
|
|
|
|
|
read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
|
|
|
|
for (tier = 1; tier < MAX_NR_TIERS; tier++) {
|
|
|
|
read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
|
|
|
|
if (!positive_ctrl_err(&sp, &pv))
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
*tier_idx = tier - 1;
|
|
|
|
|
|
|
|
return type;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
|
|
|
|
int *type_scanned, struct list_head *list)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
int type;
|
|
|
|
int scanned;
|
|
|
|
int tier = -1;
|
|
|
|
DEFINE_MIN_SEQ(lruvec);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Try to make the obvious choice first. When anon and file are both
|
|
|
|
* available from the same generation, interpret swappiness 1 as file
|
|
|
|
* first and 200 as anon first.
|
|
|
|
*/
|
|
|
|
if (!swappiness)
|
|
|
|
type = LRU_GEN_FILE;
|
|
|
|
else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
|
|
|
|
type = LRU_GEN_ANON;
|
|
|
|
else if (swappiness == 1)
|
|
|
|
type = LRU_GEN_FILE;
|
|
|
|
else if (swappiness == 200)
|
|
|
|
type = LRU_GEN_ANON;
|
|
|
|
else
|
|
|
|
type = get_type_to_scan(lruvec, swappiness, &tier);
|
|
|
|
|
|
|
|
for (i = !swappiness; i < ANON_AND_FILE; i++) {
|
|
|
|
if (tier < 0)
|
|
|
|
tier = get_tier_idx(lruvec, type);
|
|
|
|
|
|
|
|
scanned = scan_folios(lruvec, sc, type, tier, list);
|
|
|
|
if (scanned)
|
|
|
|
break;
|
|
|
|
|
|
|
|
type = !type;
|
|
|
|
tier = -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
*type_scanned = type;
|
|
|
|
|
|
|
|
return scanned;
|
|
|
|
}
|
|
|
|
|
2022-12-22 04:19:01 +00:00
|
|
|
static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness)
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
{
|
|
|
|
int type;
|
|
|
|
int scanned;
|
|
|
|
int reclaimed;
|
|
|
|
LIST_HEAD(list);
|
mm: multi-gen LRU: retry folios written back while isolated
The page reclaim isolates a batch of folios from the tail of one of the
LRU lists and works on those folios one by one. For a suitable
swap-backed folio, if the swap device is async, it queues that folio for
writeback. After the page reclaim finishes an entire batch, it puts back
the folios it queued for writeback to the head of the original LRU list.
In the meantime, the page writeback flushes the queued folios also by
batches. Its batching logic is independent from that of the page reclaim.
For each of the folios it writes back, the page writeback calls
folio_rotate_reclaimable() which tries to rotate a folio to the tail.
folio_rotate_reclaimable() only works for a folio after the page reclaim
has put it back. If an async swap device is fast enough, the page
writeback can finish with that folio while the page reclaim is still
working on the rest of the batch containing it. In this case, that folio
will remain at the head and the page reclaim will not retry it before
reaching there.
This patch adds a retry to evict_folios(). After evict_folios() has
finished an entire batch and before it puts back folios it cannot free
immediately, it retries those that may have missed the rotation.
Before this patch, ~60% of folios swapped to an Intel Optane missed
folio_rotate_reclaimable(). After this patch, ~99% of missed folios were
reclaimed upon retry.
This problem affects relatively slow async swap devices like Samsung 980
Pro much less and does not affect sync swap devices like zram or zswap at
all.
Link: https://lkml.kernel.org/r/20221116013808.3995280-1-yuzhao@google.com
Fixes: ac35a4902374 ("mm: multi-gen LRU: minimal implementation")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 01:38:07 +00:00
|
|
|
LIST_HEAD(clean);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
struct folio *folio;
|
mm: multi-gen LRU: retry folios written back while isolated
The page reclaim isolates a batch of folios from the tail of one of the
LRU lists and works on those folios one by one. For a suitable
swap-backed folio, if the swap device is async, it queues that folio for
writeback. After the page reclaim finishes an entire batch, it puts back
the folios it queued for writeback to the head of the original LRU list.
In the meantime, the page writeback flushes the queued folios also by
batches. Its batching logic is independent from that of the page reclaim.
For each of the folios it writes back, the page writeback calls
folio_rotate_reclaimable() which tries to rotate a folio to the tail.
folio_rotate_reclaimable() only works for a folio after the page reclaim
has put it back. If an async swap device is fast enough, the page
writeback can finish with that folio while the page reclaim is still
working on the rest of the batch containing it. In this case, that folio
will remain at the head and the page reclaim will not retry it before
reaching there.
This patch adds a retry to evict_folios(). After evict_folios() has
finished an entire batch and before it puts back folios it cannot free
immediately, it retries those that may have missed the rotation.
Before this patch, ~60% of folios swapped to an Intel Optane missed
folio_rotate_reclaimable(). After this patch, ~99% of missed folios were
reclaimed upon retry.
This problem affects relatively slow async swap devices like Samsung 980
Pro much less and does not affect sync swap devices like zram or zswap at
all.
Link: https://lkml.kernel.org/r/20221116013808.3995280-1-yuzhao@google.com
Fixes: ac35a4902374 ("mm: multi-gen LRU: minimal implementation")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 01:38:07 +00:00
|
|
|
struct folio *next;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
enum vm_event_item item;
|
|
|
|
struct reclaim_stat stat;
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
struct lru_gen_mm_walk *walk;
|
mm: multi-gen LRU: retry folios written back while isolated
The page reclaim isolates a batch of folios from the tail of one of the
LRU lists and works on those folios one by one. For a suitable
swap-backed folio, if the swap device is async, it queues that folio for
writeback. After the page reclaim finishes an entire batch, it puts back
the folios it queued for writeback to the head of the original LRU list.
In the meantime, the page writeback flushes the queued folios also by
batches. Its batching logic is independent from that of the page reclaim.
For each of the folios it writes back, the page writeback calls
folio_rotate_reclaimable() which tries to rotate a folio to the tail.
folio_rotate_reclaimable() only works for a folio after the page reclaim
has put it back. If an async swap device is fast enough, the page
writeback can finish with that folio while the page reclaim is still
working on the rest of the batch containing it. In this case, that folio
will remain at the head and the page reclaim will not retry it before
reaching there.
This patch adds a retry to evict_folios(). After evict_folios() has
finished an entire batch and before it puts back folios it cannot free
immediately, it retries those that may have missed the rotation.
Before this patch, ~60% of folios swapped to an Intel Optane missed
folio_rotate_reclaimable(). After this patch, ~99% of missed folios were
reclaimed upon retry.
This problem affects relatively slow async swap devices like Samsung 980
Pro much less and does not affect sync swap devices like zram or zswap at
all.
Link: https://lkml.kernel.org/r/20221116013808.3995280-1-yuzhao@google.com
Fixes: ac35a4902374 ("mm: multi-gen LRU: minimal implementation")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 01:38:07 +00:00
|
|
|
bool skip_retry = false;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(lruvec);
|
|
|
|
|
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
|
|
|
|
|
|
|
scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
|
|
|
|
|
|
|
|
scanned += try_to_inc_min_seq(lruvec, swappiness);
|
|
|
|
|
|
|
|
if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
|
|
|
|
scanned = 0;
|
|
|
|
|
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
|
|
|
|
|
|
|
if (list_empty(&list))
|
|
|
|
return scanned;
|
mm: multi-gen LRU: retry folios written back while isolated
The page reclaim isolates a batch of folios from the tail of one of the
LRU lists and works on those folios one by one. For a suitable
swap-backed folio, if the swap device is async, it queues that folio for
writeback. After the page reclaim finishes an entire batch, it puts back
the folios it queued for writeback to the head of the original LRU list.
In the meantime, the page writeback flushes the queued folios also by
batches. Its batching logic is independent from that of the page reclaim.
For each of the folios it writes back, the page writeback calls
folio_rotate_reclaimable() which tries to rotate a folio to the tail.
folio_rotate_reclaimable() only works for a folio after the page reclaim
has put it back. If an async swap device is fast enough, the page
writeback can finish with that folio while the page reclaim is still
working on the rest of the batch containing it. In this case, that folio
will remain at the head and the page reclaim will not retry it before
reaching there.
This patch adds a retry to evict_folios(). After evict_folios() has
finished an entire batch and before it puts back folios it cannot free
immediately, it retries those that may have missed the rotation.
Before this patch, ~60% of folios swapped to an Intel Optane missed
folio_rotate_reclaimable(). After this patch, ~99% of missed folios were
reclaimed upon retry.
This problem affects relatively slow async swap devices like Samsung 980
Pro much less and does not affect sync swap devices like zram or zswap at
all.
Link: https://lkml.kernel.org/r/20221116013808.3995280-1-yuzhao@google.com
Fixes: ac35a4902374 ("mm: multi-gen LRU: minimal implementation")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 01:38:07 +00:00
|
|
|
retry:
|
2022-09-02 19:45:57 +00:00
|
|
|
reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
|
mm: multi-gen LRU: retry folios written back while isolated
The page reclaim isolates a batch of folios from the tail of one of the
LRU lists and works on those folios one by one. For a suitable
swap-backed folio, if the swap device is async, it queues that folio for
writeback. After the page reclaim finishes an entire batch, it puts back
the folios it queued for writeback to the head of the original LRU list.
In the meantime, the page writeback flushes the queued folios also by
batches. Its batching logic is independent from that of the page reclaim.
For each of the folios it writes back, the page writeback calls
folio_rotate_reclaimable() which tries to rotate a folio to the tail.
folio_rotate_reclaimable() only works for a folio after the page reclaim
has put it back. If an async swap device is fast enough, the page
writeback can finish with that folio while the page reclaim is still
working on the rest of the batch containing it. In this case, that folio
will remain at the head and the page reclaim will not retry it before
reaching there.
This patch adds a retry to evict_folios(). After evict_folios() has
finished an entire batch and before it puts back folios it cannot free
immediately, it retries those that may have missed the rotation.
Before this patch, ~60% of folios swapped to an Intel Optane missed
folio_rotate_reclaimable(). After this patch, ~99% of missed folios were
reclaimed upon retry.
This problem affects relatively slow async swap devices like Samsung 980
Pro much less and does not affect sync swap devices like zram or zswap at
all.
Link: https://lkml.kernel.org/r/20221116013808.3995280-1-yuzhao@google.com
Fixes: ac35a4902374 ("mm: multi-gen LRU: minimal implementation")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 01:38:07 +00:00
|
|
|
sc->nr_reclaimed += reclaimed;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
mm: multi-gen LRU: retry folios written back while isolated
The page reclaim isolates a batch of folios from the tail of one of the
LRU lists and works on those folios one by one. For a suitable
swap-backed folio, if the swap device is async, it queues that folio for
writeback. After the page reclaim finishes an entire batch, it puts back
the folios it queued for writeback to the head of the original LRU list.
In the meantime, the page writeback flushes the queued folios also by
batches. Its batching logic is independent from that of the page reclaim.
For each of the folios it writes back, the page writeback calls
folio_rotate_reclaimable() which tries to rotate a folio to the tail.
folio_rotate_reclaimable() only works for a folio after the page reclaim
has put it back. If an async swap device is fast enough, the page
writeback can finish with that folio while the page reclaim is still
working on the rest of the batch containing it. In this case, that folio
will remain at the head and the page reclaim will not retry it before
reaching there.
This patch adds a retry to evict_folios(). After evict_folios() has
finished an entire batch and before it puts back folios it cannot free
immediately, it retries those that may have missed the rotation.
Before this patch, ~60% of folios swapped to an Intel Optane missed
folio_rotate_reclaimable(). After this patch, ~99% of missed folios were
reclaimed upon retry.
This problem affects relatively slow async swap devices like Samsung 980
Pro much less and does not affect sync swap devices like zram or zswap at
all.
Link: https://lkml.kernel.org/r/20221116013808.3995280-1-yuzhao@google.com
Fixes: ac35a4902374 ("mm: multi-gen LRU: minimal implementation")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 01:38:07 +00:00
|
|
|
list_for_each_entry_safe_reverse(folio, next, &list, lru) {
|
|
|
|
if (!folio_evictable(folio)) {
|
|
|
|
list_del(&folio->lru);
|
|
|
|
folio_putback_lru(folio);
|
|
|
|
continue;
|
|
|
|
}
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
if (folio_test_reclaim(folio) &&
|
mm: multi-gen LRU: retry folios written back while isolated
The page reclaim isolates a batch of folios from the tail of one of the
LRU lists and works on those folios one by one. For a suitable
swap-backed folio, if the swap device is async, it queues that folio for
writeback. After the page reclaim finishes an entire batch, it puts back
the folios it queued for writeback to the head of the original LRU list.
In the meantime, the page writeback flushes the queued folios also by
batches. Its batching logic is independent from that of the page reclaim.
For each of the folios it writes back, the page writeback calls
folio_rotate_reclaimable() which tries to rotate a folio to the tail.
folio_rotate_reclaimable() only works for a folio after the page reclaim
has put it back. If an async swap device is fast enough, the page
writeback can finish with that folio while the page reclaim is still
working on the rest of the batch containing it. In this case, that folio
will remain at the head and the page reclaim will not retry it before
reaching there.
This patch adds a retry to evict_folios(). After evict_folios() has
finished an entire batch and before it puts back folios it cannot free
immediately, it retries those that may have missed the rotation.
Before this patch, ~60% of folios swapped to an Intel Optane missed
folio_rotate_reclaimable(). After this patch, ~99% of missed folios were
reclaimed upon retry.
This problem affects relatively slow async swap devices like Samsung 980
Pro much less and does not affect sync swap devices like zram or zswap at
all.
Link: https://lkml.kernel.org/r/20221116013808.3995280-1-yuzhao@google.com
Fixes: ac35a4902374 ("mm: multi-gen LRU: minimal implementation")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 01:38:07 +00:00
|
|
|
(folio_test_dirty(folio) || folio_test_writeback(folio))) {
|
|
|
|
/* restore LRU_REFS_FLAGS cleared by isolate_folio() */
|
|
|
|
if (folio_test_workingset(folio))
|
|
|
|
folio_set_referenced(folio);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
|
|
|
|
folio_mapped(folio) || folio_test_locked(folio) ||
|
|
|
|
folio_test_dirty(folio) || folio_test_writeback(folio)) {
|
|
|
|
/* don't add rejected folios to the oldest generation */
|
|
|
|
set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
|
|
|
|
BIT(PG_active));
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* retry folios that may have missed folio_rotate_reclaimable() */
|
|
|
|
list_move(&folio->lru, &clean);
|
|
|
|
sc->nr_scanned -= folio_nr_pages(folio);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
|
|
|
|
2022-09-02 19:45:57 +00:00
|
|
|
move_folios_to_lru(lruvec, &list);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
walk = current->reclaim_state->mm_walk;
|
|
|
|
if (walk && walk->batched)
|
|
|
|
reset_batch_size(lruvec, walk);
|
|
|
|
|
2022-10-26 18:01:33 +00:00
|
|
|
item = PGSTEAL_KSWAPD + reclaimer_offset();
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
if (!cgroup_reclaim(sc))
|
|
|
|
__count_vm_events(item, reclaimed);
|
|
|
|
__count_memcg_events(memcg, item, reclaimed);
|
|
|
|
__count_vm_events(PGSTEAL_ANON + type, reclaimed);
|
|
|
|
|
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
|
|
|
|
|
|
|
mem_cgroup_uncharge_list(&list);
|
|
|
|
free_unref_page_list(&list);
|
|
|
|
|
mm: multi-gen LRU: retry folios written back while isolated
The page reclaim isolates a batch of folios from the tail of one of the
LRU lists and works on those folios one by one. For a suitable
swap-backed folio, if the swap device is async, it queues that folio for
writeback. After the page reclaim finishes an entire batch, it puts back
the folios it queued for writeback to the head of the original LRU list.
In the meantime, the page writeback flushes the queued folios also by
batches. Its batching logic is independent from that of the page reclaim.
For each of the folios it writes back, the page writeback calls
folio_rotate_reclaimable() which tries to rotate a folio to the tail.
folio_rotate_reclaimable() only works for a folio after the page reclaim
has put it back. If an async swap device is fast enough, the page
writeback can finish with that folio while the page reclaim is still
working on the rest of the batch containing it. In this case, that folio
will remain at the head and the page reclaim will not retry it before
reaching there.
This patch adds a retry to evict_folios(). After evict_folios() has
finished an entire batch and before it puts back folios it cannot free
immediately, it retries those that may have missed the rotation.
Before this patch, ~60% of folios swapped to an Intel Optane missed
folio_rotate_reclaimable(). After this patch, ~99% of missed folios were
reclaimed upon retry.
This problem affects relatively slow async swap devices like Samsung 980
Pro much less and does not affect sync swap devices like zram or zswap at
all.
Link: https://lkml.kernel.org/r/20221116013808.3995280-1-yuzhao@google.com
Fixes: ac35a4902374 ("mm: multi-gen LRU: minimal implementation")
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 01:38:07 +00:00
|
|
|
INIT_LIST_HEAD(&list);
|
|
|
|
list_splice_init(&clean, &list);
|
|
|
|
|
|
|
|
if (!list_empty(&list)) {
|
|
|
|
skip_retry = true;
|
|
|
|
goto retry;
|
|
|
|
}
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
return scanned;
|
|
|
|
}
|
|
|
|
|
2022-12-22 04:19:03 +00:00
|
|
|
static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq,
|
|
|
|
struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan)
|
|
|
|
{
|
|
|
|
int gen, type, zone;
|
|
|
|
unsigned long old = 0;
|
|
|
|
unsigned long young = 0;
|
|
|
|
unsigned long total = 0;
|
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
DEFINE_MIN_SEQ(lruvec);
|
|
|
|
|
|
|
|
/* whether this lruvec is completely out of cold folios */
|
|
|
|
if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) {
|
|
|
|
*nr_to_scan = 0;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (type = !can_swap; type < ANON_AND_FILE; type++) {
|
|
|
|
unsigned long seq;
|
|
|
|
|
|
|
|
for (seq = min_seq[type]; seq <= max_seq; seq++) {
|
|
|
|
unsigned long size = 0;
|
|
|
|
|
|
|
|
gen = lru_gen_from_seq(seq);
|
|
|
|
|
|
|
|
for (zone = 0; zone < MAX_NR_ZONES; zone++)
|
|
|
|
size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
|
|
|
|
|
|
|
|
total += size;
|
|
|
|
if (seq == max_seq)
|
|
|
|
young += size;
|
|
|
|
else if (seq + MIN_NR_GENS == max_seq)
|
|
|
|
old += size;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* try to scrape all its memory if this memcg was deleted */
|
|
|
|
*nr_to_scan = mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The aging tries to be lazy to reduce the overhead, while the eviction
|
|
|
|
* stalls when the number of generations reaches MIN_NR_GENS. Hence, the
|
|
|
|
* ideal number of generations is MIN_NR_GENS+1.
|
|
|
|
*/
|
|
|
|
if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
|
|
|
|
* of the total number of pages for each generation. A reasonable range
|
|
|
|
* for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
|
|
|
|
* aging cares about the upper bound of hot pages, while the eviction
|
|
|
|
* cares about the lower bound of cold pages.
|
|
|
|
*/
|
|
|
|
if (young * MIN_NR_GENS > total)
|
|
|
|
return true;
|
|
|
|
if (old * (MIN_NR_GENS + 2) < total)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
/*
|
|
|
|
* For future optimizations:
|
|
|
|
* 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
|
|
|
|
* reclaim.
|
|
|
|
*/
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
static long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, bool can_swap)
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
{
|
|
|
|
unsigned long nr_to_scan;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
DEFINE_MAX_SEQ(lruvec);
|
|
|
|
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg))
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
return 0;
|
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
if (!should_run_aging(lruvec, max_seq, sc, can_swap, &nr_to_scan))
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
return nr_to_scan;
|
|
|
|
|
|
|
|
/* skip the aging path at the default priority */
|
|
|
|
if (sc->priority == DEF_PRIORITY)
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
return nr_to_scan;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
/* skip this lruvec as it's low on cold folios */
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
return try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false) ? -1 : 0;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
}
|
|
|
|
|
2022-12-22 04:19:01 +00:00
|
|
|
static unsigned long get_nr_to_reclaim(struct scan_control *sc)
|
mm: multi-gen LRU: optimize multiple memcgs
When multiple memcgs are available, it is possible to use generations as a
frame of reference to make better choices and improve overall performance
under global memory pressure. This patch adds a basic optimization to
select memcgs that can drop single-use unmapped clean pages first. Doing
so reduces the chance of going into the aging path or swapping, which can
be costly.
A typical example that benefits from this optimization is a server running
mixed types of workloads, e.g., heavy anon workload in one memcg and heavy
buffered I/O workload in the other.
Though this optimization can be applied to both kswapd and direct reclaim,
it is only added to kswapd to keep the patchset manageable. Later
improvements may cover the direct reclaim path.
While ensuring certain fairness to all eligible memcgs, proportional scans
of individual memcgs also require proper backoff to avoid overshooting
their aggregate reclaim target by too much. Otherwise it can cause high
direct reclaim latency. The conditions for backoff are:
1. At low priorities, for direct reclaim, if aging fairness or direct
reclaim latency is at risk, i.e., aging one memcg multiple times or
swapping after the target is met.
2. At high priorities, for global reclaim, if per-zone free pages are
above respective watermarks.
Server benchmark results:
Mixed workloads:
fio (buffered I/O): +[19, 21]%
IOPS BW
patch1-8: 1880k 7343MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[119, 123]%
Ops/sec KB/sec
patch1-8: 862768.65 33514.68
patch1-9: 1911022.12 74234.54
Mixed workloads:
fio (buffered I/O): +[75, 77]%
IOPS BW
5.19-rc1: 1279k 4996MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[13, 15]%
Ops/sec KB/sec
5.19-rc1: 1673524.04 65008.87
patch1-9: 1911022.12 74234.54
Configurations:
(changes since patch 6)
cat mixed.sh
modprobe brd rd_nr=2 rd_size=56623104
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
mkfs.ext4 /dev/ram1
mount -t ext4 /dev/ram1 /mnt
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
fio -name=mglru --numjobs=36 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=90m --group_reporting &
pid=$!
sleep 200
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
kill -INT $pid
wait
Client benchmark results:
no change (CONFIG_MEMCG=n)
Link: https://lkml.kernel.org/r/20220918080010.2920238-10-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:06 +00:00
|
|
|
{
|
2022-12-22 04:19:01 +00:00
|
|
|
/* don't abort memcg reclaim to ensure fairness */
|
|
|
|
if (!global_reclaim(sc))
|
|
|
|
return -1;
|
mm: multi-gen LRU: optimize multiple memcgs
When multiple memcgs are available, it is possible to use generations as a
frame of reference to make better choices and improve overall performance
under global memory pressure. This patch adds a basic optimization to
select memcgs that can drop single-use unmapped clean pages first. Doing
so reduces the chance of going into the aging path or swapping, which can
be costly.
A typical example that benefits from this optimization is a server running
mixed types of workloads, e.g., heavy anon workload in one memcg and heavy
buffered I/O workload in the other.
Though this optimization can be applied to both kswapd and direct reclaim,
it is only added to kswapd to keep the patchset manageable. Later
improvements may cover the direct reclaim path.
While ensuring certain fairness to all eligible memcgs, proportional scans
of individual memcgs also require proper backoff to avoid overshooting
their aggregate reclaim target by too much. Otherwise it can cause high
direct reclaim latency. The conditions for backoff are:
1. At low priorities, for direct reclaim, if aging fairness or direct
reclaim latency is at risk, i.e., aging one memcg multiple times or
swapping after the target is met.
2. At high priorities, for global reclaim, if per-zone free pages are
above respective watermarks.
Server benchmark results:
Mixed workloads:
fio (buffered I/O): +[19, 21]%
IOPS BW
patch1-8: 1880k 7343MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[119, 123]%
Ops/sec KB/sec
patch1-8: 862768.65 33514.68
patch1-9: 1911022.12 74234.54
Mixed workloads:
fio (buffered I/O): +[75, 77]%
IOPS BW
5.19-rc1: 1279k 4996MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[13, 15]%
Ops/sec KB/sec
5.19-rc1: 1673524.04 65008.87
patch1-9: 1911022.12 74234.54
Configurations:
(changes since patch 6)
cat mixed.sh
modprobe brd rd_nr=2 rd_size=56623104
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
mkfs.ext4 /dev/ram1
mount -t ext4 /dev/ram1 /mnt
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
fio -name=mglru --numjobs=36 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=90m --group_reporting &
pid=$!
sleep 200
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
kill -INT $pid
wait
Client benchmark results:
no change (CONFIG_MEMCG=n)
Link: https://lkml.kernel.org/r/20220918080010.2920238-10-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:06 +00:00
|
|
|
|
2022-12-22 04:19:01 +00:00
|
|
|
return max(sc->nr_to_reclaim, compact_gap(sc->order));
|
mm: multi-gen LRU: optimize multiple memcgs
When multiple memcgs are available, it is possible to use generations as a
frame of reference to make better choices and improve overall performance
under global memory pressure. This patch adds a basic optimization to
select memcgs that can drop single-use unmapped clean pages first. Doing
so reduces the chance of going into the aging path or swapping, which can
be costly.
A typical example that benefits from this optimization is a server running
mixed types of workloads, e.g., heavy anon workload in one memcg and heavy
buffered I/O workload in the other.
Though this optimization can be applied to both kswapd and direct reclaim,
it is only added to kswapd to keep the patchset manageable. Later
improvements may cover the direct reclaim path.
While ensuring certain fairness to all eligible memcgs, proportional scans
of individual memcgs also require proper backoff to avoid overshooting
their aggregate reclaim target by too much. Otherwise it can cause high
direct reclaim latency. The conditions for backoff are:
1. At low priorities, for direct reclaim, if aging fairness or direct
reclaim latency is at risk, i.e., aging one memcg multiple times or
swapping after the target is met.
2. At high priorities, for global reclaim, if per-zone free pages are
above respective watermarks.
Server benchmark results:
Mixed workloads:
fio (buffered I/O): +[19, 21]%
IOPS BW
patch1-8: 1880k 7343MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[119, 123]%
Ops/sec KB/sec
patch1-8: 862768.65 33514.68
patch1-9: 1911022.12 74234.54
Mixed workloads:
fio (buffered I/O): +[75, 77]%
IOPS BW
5.19-rc1: 1279k 4996MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[13, 15]%
Ops/sec KB/sec
5.19-rc1: 1673524.04 65008.87
patch1-9: 1911022.12 74234.54
Configurations:
(changes since patch 6)
cat mixed.sh
modprobe brd rd_nr=2 rd_size=56623104
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
mkfs.ext4 /dev/ram1
mount -t ext4 /dev/ram1 /mnt
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
fio -name=mglru --numjobs=36 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=90m --group_reporting &
pid=$!
sleep 200
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
kill -INT $pid
wait
Client benchmark results:
no change (CONFIG_MEMCG=n)
Link: https://lkml.kernel.org/r/20220918080010.2920238-10-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:06 +00:00
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
static bool try_to_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
{
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
long nr_to_scan;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
unsigned long scanned = 0;
|
2022-12-22 04:19:01 +00:00
|
|
|
unsigned long nr_to_reclaim = get_nr_to_reclaim(sc);
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
int swappiness = get_swappiness(lruvec, sc);
|
|
|
|
|
|
|
|
/* clean file folios are more likely to exist */
|
|
|
|
if (swappiness && !(sc->gfp_mask & __GFP_IO))
|
|
|
|
swappiness = 1;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
while (true) {
|
|
|
|
int delta;
|
|
|
|
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness);
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
if (nr_to_scan <= 0)
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
break;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
2022-12-22 04:19:01 +00:00
|
|
|
delta = evict_folios(lruvec, sc, swappiness);
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
if (!delta)
|
mm: multi-gen LRU: remove aging fairness safeguard
Recall that the aging produces the youngest generation: first it scans
for accessed folios and updates their gen counters; then it increments
lrugen->max_seq.
The current aging fairness safeguard for kswapd uses two passes to
ensure the fairness to multiple eligible memcgs. On the first pass,
which is shared with the eviction, it checks whether all eligible
memcgs are low on cold folios. If so, it requires a second pass, on
which it ages all those memcgs at the same time.
With memcg LRU, the aging, while ensuring eventual fairness, will run
when necessary. Therefore the current aging fairness safeguard for
kswapd will not be needed.
Note that memcg LRU only applies to global reclaim. For memcg reclaim,
the aging can be unfair to different memcgs, i.e., their
lrugen->max_seq can be incremented at different paces.
Link: https://lkml.kernel.org/r/20221222041905.2431096-5-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:02 +00:00
|
|
|
break;
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
|
|
|
|
scanned += delta;
|
|
|
|
if (scanned >= nr_to_scan)
|
|
|
|
break;
|
|
|
|
|
2022-12-22 04:19:01 +00:00
|
|
|
if (sc->nr_reclaimed >= nr_to_reclaim)
|
mm: multi-gen LRU: optimize multiple memcgs
When multiple memcgs are available, it is possible to use generations as a
frame of reference to make better choices and improve overall performance
under global memory pressure. This patch adds a basic optimization to
select memcgs that can drop single-use unmapped clean pages first. Doing
so reduces the chance of going into the aging path or swapping, which can
be costly.
A typical example that benefits from this optimization is a server running
mixed types of workloads, e.g., heavy anon workload in one memcg and heavy
buffered I/O workload in the other.
Though this optimization can be applied to both kswapd and direct reclaim,
it is only added to kswapd to keep the patchset manageable. Later
improvements may cover the direct reclaim path.
While ensuring certain fairness to all eligible memcgs, proportional scans
of individual memcgs also require proper backoff to avoid overshooting
their aggregate reclaim target by too much. Otherwise it can cause high
direct reclaim latency. The conditions for backoff are:
1. At low priorities, for direct reclaim, if aging fairness or direct
reclaim latency is at risk, i.e., aging one memcg multiple times or
swapping after the target is met.
2. At high priorities, for global reclaim, if per-zone free pages are
above respective watermarks.
Server benchmark results:
Mixed workloads:
fio (buffered I/O): +[19, 21]%
IOPS BW
patch1-8: 1880k 7343MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[119, 123]%
Ops/sec KB/sec
patch1-8: 862768.65 33514.68
patch1-9: 1911022.12 74234.54
Mixed workloads:
fio (buffered I/O): +[75, 77]%
IOPS BW
5.19-rc1: 1279k 4996MiB/s
patch1-9: 2252k 8796MiB/s
memcached (anon): +[13, 15]%
Ops/sec KB/sec
5.19-rc1: 1673524.04 65008.87
patch1-9: 1911022.12 74234.54
Configurations:
(changes since patch 6)
cat mixed.sh
modprobe brd rd_nr=2 rd_size=56623104
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
mkfs.ext4 /dev/ram1
mount -t ext4 /dev/ram1 /mnt
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
fio -name=mglru --numjobs=36 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=90m --group_reporting &
pid=$!
sleep 200
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=50000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
kill -INT $pid
wait
Client benchmark results:
no change (CONFIG_MEMCG=n)
Link: https://lkml.kernel.org/r/20220918080010.2920238-10-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:06 +00:00
|
|
|
break;
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
cond_resched();
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
/* whether try_to_inc_max_seq() was successful */
|
|
|
|
return nr_to_scan < 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int shrink_one(struct lruvec *lruvec, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
bool success;
|
|
|
|
unsigned long scanned = sc->nr_scanned;
|
|
|
|
unsigned long reclaimed = sc->nr_reclaimed;
|
|
|
|
int seg = lru_gen_memcg_seg(lruvec);
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
struct pglist_data *pgdat = lruvec_pgdat(lruvec);
|
|
|
|
|
|
|
|
/* see the comment on MEMCG_NR_GENS */
|
|
|
|
if (!lruvec_is_sizable(lruvec, sc))
|
|
|
|
return seg != MEMCG_LRU_TAIL ? MEMCG_LRU_TAIL : MEMCG_LRU_YOUNG;
|
|
|
|
|
|
|
|
mem_cgroup_calculate_protection(NULL, memcg);
|
|
|
|
|
|
|
|
if (mem_cgroup_below_min(NULL, memcg))
|
|
|
|
return MEMCG_LRU_YOUNG;
|
|
|
|
|
|
|
|
if (mem_cgroup_below_low(NULL, memcg)) {
|
|
|
|
/* see the comment on MEMCG_NR_GENS */
|
|
|
|
if (seg != MEMCG_LRU_TAIL)
|
|
|
|
return MEMCG_LRU_TAIL;
|
|
|
|
|
|
|
|
memcg_memory_event(memcg, MEMCG_LOW);
|
|
|
|
}
|
|
|
|
|
|
|
|
success = try_to_shrink_lruvec(lruvec, sc);
|
|
|
|
|
|
|
|
shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, sc->priority);
|
|
|
|
|
|
|
|
if (!sc->proactive)
|
|
|
|
vmpressure(sc->gfp_mask, memcg, false, sc->nr_scanned - scanned,
|
|
|
|
sc->nr_reclaimed - reclaimed);
|
|
|
|
|
|
|
|
sc->nr_reclaimed += current->reclaim_state->reclaimed_slab;
|
|
|
|
current->reclaim_state->reclaimed_slab = 0;
|
|
|
|
|
|
|
|
return success ? MEMCG_LRU_YOUNG : 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
|
|
|
|
static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
|
|
|
|
{
|
2023-02-13 07:53:22 +00:00
|
|
|
int op;
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
int gen;
|
|
|
|
int bin;
|
|
|
|
int first_bin;
|
|
|
|
struct lruvec *lruvec;
|
|
|
|
struct lru_gen_folio *lrugen;
|
2023-02-13 07:53:22 +00:00
|
|
|
struct mem_cgroup *memcg;
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
const struct hlist_nulls_node *pos;
|
|
|
|
unsigned long nr_to_reclaim = get_nr_to_reclaim(sc);
|
|
|
|
|
|
|
|
bin = first_bin = get_random_u32_below(MEMCG_NR_BINS);
|
|
|
|
restart:
|
2023-02-13 07:53:22 +00:00
|
|
|
op = 0;
|
|
|
|
memcg = NULL;
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
gen = get_memcg_gen(READ_ONCE(pgdat->memcg_lru.seq));
|
|
|
|
|
|
|
|
rcu_read_lock();
|
|
|
|
|
|
|
|
hlist_nulls_for_each_entry_rcu(lrugen, pos, &pgdat->memcg_lru.fifo[gen][bin], list) {
|
|
|
|
if (op)
|
|
|
|
lru_gen_rotate_memcg(lruvec, op);
|
|
|
|
|
|
|
|
mem_cgroup_put(memcg);
|
|
|
|
|
|
|
|
lruvec = container_of(lrugen, struct lruvec, lrugen);
|
|
|
|
memcg = lruvec_memcg(lruvec);
|
|
|
|
|
|
|
|
if (!mem_cgroup_tryget(memcg)) {
|
|
|
|
op = 0;
|
|
|
|
memcg = NULL;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
rcu_read_unlock();
|
|
|
|
|
|
|
|
op = shrink_one(lruvec, sc);
|
|
|
|
|
|
|
|
rcu_read_lock();
|
2023-02-13 07:53:22 +00:00
|
|
|
|
|
|
|
if (sc->nr_reclaimed >= nr_to_reclaim)
|
|
|
|
break;
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
rcu_read_unlock();
|
|
|
|
|
2023-02-13 07:53:22 +00:00
|
|
|
if (op)
|
|
|
|
lru_gen_rotate_memcg(lruvec, op);
|
|
|
|
|
|
|
|
mem_cgroup_put(memcg);
|
|
|
|
|
|
|
|
if (sc->nr_reclaimed >= nr_to_reclaim)
|
|
|
|
return;
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
/* restart if raced with lru_gen_rotate_memcg() */
|
|
|
|
if (gen != get_nulls_value(pos))
|
|
|
|
goto restart;
|
|
|
|
|
|
|
|
/* try the rest of the bins of the current generation */
|
|
|
|
bin = get_memcg_bin(bin + 1);
|
|
|
|
if (bin != first_bin)
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
struct blk_plug plug;
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(global_reclaim(sc));
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
VM_WARN_ON_ONCE(!sc->may_writepage || !sc->may_unmap);
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
|
|
|
|
lru_add_drain();
|
|
|
|
|
|
|
|
blk_start_plug(&plug);
|
|
|
|
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
set_mm_walk(NULL, sc->proactive);
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
|
|
|
|
if (try_to_shrink_lruvec(lruvec, sc))
|
|
|
|
lru_gen_rotate_memcg(lruvec, MEMCG_LRU_YOUNG);
|
|
|
|
|
|
|
|
clear_mm_walk();
|
|
|
|
|
|
|
|
blk_finish_plug(&plug);
|
|
|
|
}
|
|
|
|
|
|
|
|
#else /* !CONFIG_MEMCG */
|
|
|
|
|
|
|
|
static void shrink_many(struct pglist_data *pgdat, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
BUILD_BUG();
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
BUILD_BUG();
|
|
|
|
}
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
static void set_initial_priority(struct pglist_data *pgdat, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
int priority;
|
|
|
|
unsigned long reclaimable;
|
|
|
|
struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
|
|
|
|
|
|
|
|
if (sc->priority != DEF_PRIORITY || sc->nr_to_reclaim < MIN_LRU_BATCH)
|
|
|
|
return;
|
|
|
|
/*
|
|
|
|
* Determine the initial priority based on ((total / MEMCG_NR_GENS) >>
|
|
|
|
* priority) * reclaimed_to_scanned_ratio = nr_to_reclaim, where the
|
|
|
|
* estimated reclaimed_to_scanned_ratio = inactive / total.
|
|
|
|
*/
|
|
|
|
reclaimable = node_page_state(pgdat, NR_INACTIVE_FILE);
|
|
|
|
if (get_swappiness(lruvec, sc))
|
|
|
|
reclaimable += node_page_state(pgdat, NR_INACTIVE_ANON);
|
|
|
|
|
|
|
|
reclaimable /= MEMCG_NR_GENS;
|
|
|
|
|
|
|
|
/* round down reclaimable and round up sc->nr_to_reclaim */
|
|
|
|
priority = fls_long(reclaimable) - 1 - fls_long(sc->nr_to_reclaim - 1);
|
|
|
|
|
|
|
|
sc->priority = clamp(priority, 0, DEF_PRIORITY);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
struct blk_plug plug;
|
|
|
|
unsigned long reclaimed = sc->nr_reclaimed;
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(!global_reclaim(sc));
|
|
|
|
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
/*
|
|
|
|
* Unmapped clean folios are already prioritized. Scanning for more of
|
|
|
|
* them is likely futile and can cause high reclaim latency when there
|
|
|
|
* is a large number of memcgs.
|
|
|
|
*/
|
|
|
|
if (!sc->may_writepage || !sc->may_unmap)
|
|
|
|
goto done;
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
lru_add_drain();
|
|
|
|
|
|
|
|
blk_start_plug(&plug);
|
|
|
|
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
set_mm_walk(pgdat, sc->proactive);
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
|
|
|
|
set_initial_priority(pgdat, sc);
|
|
|
|
|
|
|
|
if (current_is_kswapd())
|
|
|
|
sc->nr_reclaimed = 0;
|
|
|
|
|
|
|
|
if (mem_cgroup_disabled())
|
|
|
|
shrink_one(&pgdat->__lruvec, sc);
|
|
|
|
else
|
|
|
|
shrink_many(pgdat, sc);
|
|
|
|
|
|
|
|
if (current_is_kswapd())
|
|
|
|
sc->nr_reclaimed += reclaimed;
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
clear_mm_walk();
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
blk_finish_plug(&plug);
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
done:
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
/* kswapd should never fail */
|
|
|
|
pgdat->kswapd_failures = 0;
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* state change
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
|
|
|
|
{
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
|
|
|
|
if (lrugen->enabled) {
|
|
|
|
enum lru_list lru;
|
|
|
|
|
|
|
|
for_each_evictable_lru(lru) {
|
|
|
|
if (!list_empty(&lruvec->lists[lru]))
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
int gen, type, zone;
|
|
|
|
|
|
|
|
for_each_gen_type_zone(gen, type, zone) {
|
2022-12-22 04:19:00 +00:00
|
|
|
if (!list_empty(&lrugen->folios[gen][type][zone]))
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool fill_evictable(struct lruvec *lruvec)
|
|
|
|
{
|
|
|
|
enum lru_list lru;
|
|
|
|
int remaining = MAX_LRU_BATCH;
|
|
|
|
|
|
|
|
for_each_evictable_lru(lru) {
|
|
|
|
int type = is_file_lru(lru);
|
|
|
|
bool active = is_active_lru(lru);
|
|
|
|
struct list_head *head = &lruvec->lists[lru];
|
|
|
|
|
|
|
|
while (!list_empty(head)) {
|
|
|
|
bool success;
|
|
|
|
struct folio *folio = lru_to_folio(head);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
|
|
|
|
|
|
|
|
lruvec_del_folio(lruvec, folio);
|
|
|
|
success = lru_gen_add_folio(lruvec, folio, false);
|
|
|
|
VM_WARN_ON_ONCE(!success);
|
|
|
|
|
|
|
|
if (!--remaining)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drain_evictable(struct lruvec *lruvec)
|
|
|
|
{
|
|
|
|
int gen, type, zone;
|
|
|
|
int remaining = MAX_LRU_BATCH;
|
|
|
|
|
|
|
|
for_each_gen_type_zone(gen, type, zone) {
|
2022-12-22 04:19:00 +00:00
|
|
|
struct list_head *head = &lruvec->lrugen.folios[gen][type][zone];
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
|
|
|
|
while (!list_empty(head)) {
|
|
|
|
bool success;
|
|
|
|
struct folio *folio = lru_to_folio(head);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
|
|
|
|
VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
|
|
|
|
|
|
|
|
success = lru_gen_del_folio(lruvec, folio, false);
|
|
|
|
VM_WARN_ON_ONCE(!success);
|
|
|
|
lruvec_add_folio(lruvec, folio);
|
|
|
|
|
|
|
|
if (!--remaining)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lru_gen_change_state(bool enabled)
|
|
|
|
{
|
|
|
|
static DEFINE_MUTEX(state_mutex);
|
|
|
|
|
|
|
|
struct mem_cgroup *memcg;
|
|
|
|
|
|
|
|
cgroup_lock();
|
|
|
|
cpus_read_lock();
|
|
|
|
get_online_mems();
|
|
|
|
mutex_lock(&state_mutex);
|
|
|
|
|
|
|
|
if (enabled == lru_gen_enabled())
|
|
|
|
goto unlock;
|
|
|
|
|
|
|
|
if (enabled)
|
|
|
|
static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
|
|
|
|
else
|
|
|
|
static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
|
|
|
|
|
|
|
|
memcg = mem_cgroup_iter(NULL, NULL, NULL);
|
|
|
|
do {
|
|
|
|
int nid;
|
|
|
|
|
|
|
|
for_each_node(nid) {
|
|
|
|
struct lruvec *lruvec = get_lruvec(memcg, nid);
|
|
|
|
|
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
|
|
|
|
|
|
|
VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
|
|
|
|
VM_WARN_ON_ONCE(!state_is_valid(lruvec));
|
|
|
|
|
|
|
|
lruvec->lrugen.enabled = enabled;
|
|
|
|
|
|
|
|
while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
|
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
|
|
|
cond_resched();
|
|
|
|
spin_lock_irq(&lruvec->lru_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
spin_unlock_irq(&lruvec->lru_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
cond_resched();
|
|
|
|
} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
|
|
|
|
unlock:
|
|
|
|
mutex_unlock(&state_mutex);
|
|
|
|
put_online_mems();
|
|
|
|
cpus_read_unlock();
|
|
|
|
cgroup_unlock();
|
|
|
|
}
|
|
|
|
|
|
|
|
/******************************************************************************
|
|
|
|
* sysfs interface
|
|
|
|
******************************************************************************/
|
|
|
|
|
2022-09-18 08:00:08 +00:00
|
|
|
static ssize_t show_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
|
|
|
|
{
|
|
|
|
return sprintf(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
|
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:10 +00:00
|
|
|
/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
|
2022-09-18 08:00:08 +00:00
|
|
|
static ssize_t store_min_ttl(struct kobject *kobj, struct kobj_attribute *attr,
|
|
|
|
const char *buf, size_t len)
|
|
|
|
{
|
|
|
|
unsigned int msecs;
|
|
|
|
|
|
|
|
if (kstrtouint(buf, 0, &msecs))
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
|
|
|
|
|
|
|
|
return len;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR(
|
|
|
|
min_ttl_ms, 0644, show_min_ttl, store_min_ttl
|
|
|
|
);
|
|
|
|
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
static ssize_t show_enabled(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
|
|
|
|
{
|
|
|
|
unsigned int caps = 0;
|
|
|
|
|
|
|
|
if (get_cap(LRU_GEN_CORE))
|
|
|
|
caps |= BIT(LRU_GEN_CORE);
|
|
|
|
|
|
|
|
if (arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))
|
|
|
|
caps |= BIT(LRU_GEN_MM_WALK);
|
|
|
|
|
2022-11-23 06:45:10 +00:00
|
|
|
if (arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG))
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
|
|
|
|
|
2022-11-24 11:29:01 +00:00
|
|
|
return sysfs_emit(buf, "0x%04x\n", caps);
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:10 +00:00
|
|
|
/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
static ssize_t store_enabled(struct kobject *kobj, struct kobj_attribute *attr,
|
|
|
|
const char *buf, size_t len)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
unsigned int caps;
|
|
|
|
|
|
|
|
if (tolower(*buf) == 'n')
|
|
|
|
caps = 0;
|
|
|
|
else if (tolower(*buf) == 'y')
|
|
|
|
caps = -1;
|
|
|
|
else if (kstrtouint(buf, 0, &caps))
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
|
|
|
|
bool enabled = caps & BIT(i);
|
|
|
|
|
|
|
|
if (i == LRU_GEN_CORE)
|
|
|
|
lru_gen_change_state(enabled);
|
|
|
|
else if (enabled)
|
|
|
|
static_branch_enable(&lru_gen_caps[i]);
|
|
|
|
else
|
|
|
|
static_branch_disable(&lru_gen_caps[i]);
|
|
|
|
}
|
|
|
|
|
|
|
|
return len;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct kobj_attribute lru_gen_enabled_attr = __ATTR(
|
|
|
|
enabled, 0644, show_enabled, store_enabled
|
|
|
|
);
|
|
|
|
|
|
|
|
static struct attribute *lru_gen_attrs[] = {
|
2022-09-18 08:00:08 +00:00
|
|
|
&lru_gen_min_ttl_attr.attr,
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
&lru_gen_enabled_attr.attr,
|
|
|
|
NULL
|
|
|
|
};
|
|
|
|
|
|
|
|
static struct attribute_group lru_gen_attr_group = {
|
|
|
|
.name = "lru_gen",
|
|
|
|
.attrs = lru_gen_attrs,
|
|
|
|
};
|
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* debugfs interface
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
|
|
|
|
{
|
|
|
|
struct mem_cgroup *memcg;
|
|
|
|
loff_t nr_to_skip = *pos;
|
|
|
|
|
|
|
|
m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
|
|
|
|
if (!m->private)
|
|
|
|
return ERR_PTR(-ENOMEM);
|
|
|
|
|
|
|
|
memcg = mem_cgroup_iter(NULL, NULL, NULL);
|
|
|
|
do {
|
|
|
|
int nid;
|
|
|
|
|
|
|
|
for_each_node_state(nid, N_MEMORY) {
|
|
|
|
if (!nr_to_skip--)
|
|
|
|
return get_lruvec(memcg, nid);
|
|
|
|
}
|
|
|
|
} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lru_gen_seq_stop(struct seq_file *m, void *v)
|
|
|
|
{
|
|
|
|
if (!IS_ERR_OR_NULL(v))
|
|
|
|
mem_cgroup_iter_break(NULL, lruvec_memcg(v));
|
|
|
|
|
|
|
|
kvfree(m->private);
|
|
|
|
m->private = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
|
|
|
|
{
|
|
|
|
int nid = lruvec_pgdat(v)->node_id;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(v);
|
|
|
|
|
|
|
|
++*pos;
|
|
|
|
|
|
|
|
nid = next_memory_node(nid);
|
|
|
|
if (nid == MAX_NUMNODES) {
|
|
|
|
memcg = mem_cgroup_iter(NULL, memcg, NULL);
|
|
|
|
if (!memcg)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
nid = first_memory_node;
|
|
|
|
}
|
|
|
|
|
|
|
|
return get_lruvec(memcg, nid);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
|
|
|
|
unsigned long max_seq, unsigned long *min_seq,
|
|
|
|
unsigned long seq)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
int type, tier;
|
|
|
|
int hist = lru_hist_from_seq(seq);
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
2022-09-18 08:00:09 +00:00
|
|
|
|
|
|
|
for (tier = 0; tier < MAX_NR_TIERS; tier++) {
|
|
|
|
seq_printf(m, " %10d", tier);
|
|
|
|
for (type = 0; type < ANON_AND_FILE; type++) {
|
|
|
|
const char *s = " ";
|
|
|
|
unsigned long n[3] = {};
|
|
|
|
|
|
|
|
if (seq == max_seq) {
|
|
|
|
s = "RT ";
|
|
|
|
n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
|
|
|
|
n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
|
|
|
|
} else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
|
|
|
|
s = "rep";
|
|
|
|
n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
|
|
|
|
n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
|
|
|
|
if (tier)
|
|
|
|
n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (i = 0; i < 3; i++)
|
|
|
|
seq_printf(m, " %10lu%c", n[i], s[i]);
|
|
|
|
}
|
|
|
|
seq_putc(m, '\n');
|
|
|
|
}
|
|
|
|
|
|
|
|
seq_puts(m, " ");
|
|
|
|
for (i = 0; i < NR_MM_STATS; i++) {
|
|
|
|
const char *s = " ";
|
|
|
|
unsigned long n = 0;
|
|
|
|
|
|
|
|
if (seq == max_seq && NR_HIST_GENS == 1) {
|
|
|
|
s = "LOYNFA";
|
|
|
|
n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
|
|
|
|
} else if (seq != max_seq && NR_HIST_GENS > 1) {
|
|
|
|
s = "loynfa";
|
|
|
|
n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
|
|
|
|
}
|
|
|
|
|
|
|
|
seq_printf(m, " %10lu%c", n, s[i]);
|
|
|
|
}
|
|
|
|
seq_putc(m, '\n');
|
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:10 +00:00
|
|
|
/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
|
2022-09-18 08:00:09 +00:00
|
|
|
static int lru_gen_seq_show(struct seq_file *m, void *v)
|
|
|
|
{
|
|
|
|
unsigned long seq;
|
|
|
|
bool full = !debugfs_real_fops(m->file)->write;
|
|
|
|
struct lruvec *lruvec = v;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
2022-09-18 08:00:09 +00:00
|
|
|
int nid = lruvec_pgdat(lruvec)->node_id;
|
|
|
|
struct mem_cgroup *memcg = lruvec_memcg(lruvec);
|
|
|
|
DEFINE_MAX_SEQ(lruvec);
|
|
|
|
DEFINE_MIN_SEQ(lruvec);
|
|
|
|
|
|
|
|
if (nid == first_memory_node) {
|
|
|
|
const char *path = memcg ? m->private : "";
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
|
|
|
if (memcg)
|
|
|
|
cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
|
|
|
|
#endif
|
|
|
|
seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
|
|
|
|
}
|
|
|
|
|
|
|
|
seq_printf(m, " node %5d\n", nid);
|
|
|
|
|
|
|
|
if (!full)
|
|
|
|
seq = min_seq[LRU_GEN_ANON];
|
|
|
|
else if (max_seq >= MAX_NR_GENS)
|
|
|
|
seq = max_seq - MAX_NR_GENS + 1;
|
|
|
|
else
|
|
|
|
seq = 0;
|
|
|
|
|
|
|
|
for (; seq <= max_seq; seq++) {
|
|
|
|
int type, zone;
|
|
|
|
int gen = lru_gen_from_seq(seq);
|
|
|
|
unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
|
|
|
|
|
|
|
|
seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
|
|
|
|
|
|
|
|
for (type = 0; type < ANON_AND_FILE; type++) {
|
|
|
|
unsigned long size = 0;
|
|
|
|
char mark = full && seq < min_seq[type] ? 'x' : ' ';
|
|
|
|
|
|
|
|
for (zone = 0; zone < MAX_NR_ZONES; zone++)
|
|
|
|
size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
|
|
|
|
|
|
|
|
seq_printf(m, " %10lu%c", size, mark);
|
|
|
|
}
|
|
|
|
|
|
|
|
seq_putc(m, '\n');
|
|
|
|
|
|
|
|
if (full)
|
|
|
|
lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static const struct seq_operations lru_gen_seq_ops = {
|
|
|
|
.start = lru_gen_seq_start,
|
|
|
|
.stop = lru_gen_seq_stop,
|
|
|
|
.next = lru_gen_seq_next,
|
|
|
|
.show = lru_gen_seq_show,
|
|
|
|
};
|
|
|
|
|
|
|
|
static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
|
|
|
|
bool can_swap, bool force_scan)
|
|
|
|
{
|
|
|
|
DEFINE_MAX_SEQ(lruvec);
|
|
|
|
DEFINE_MIN_SEQ(lruvec);
|
|
|
|
|
|
|
|
if (seq < max_seq)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (seq > max_seq)
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
|
|
|
|
return -ERANGE;
|
|
|
|
|
|
|
|
try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
|
|
|
|
int swappiness, unsigned long nr_to_reclaim)
|
|
|
|
{
|
|
|
|
DEFINE_MAX_SEQ(lruvec);
|
|
|
|
|
|
|
|
if (seq + MIN_NR_GENS > max_seq)
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
sc->nr_reclaimed = 0;
|
|
|
|
|
|
|
|
while (!signal_pending(current)) {
|
|
|
|
DEFINE_MIN_SEQ(lruvec);
|
|
|
|
|
|
|
|
if (seq < min_seq[!swappiness])
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (sc->nr_reclaimed >= nr_to_reclaim)
|
|
|
|
return 0;
|
|
|
|
|
2022-12-22 04:19:01 +00:00
|
|
|
if (!evict_folios(lruvec, sc, swappiness))
|
2022-09-18 08:00:09 +00:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
cond_resched();
|
|
|
|
}
|
|
|
|
|
|
|
|
return -EINTR;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
|
|
|
|
struct scan_control *sc, int swappiness, unsigned long opt)
|
|
|
|
{
|
|
|
|
struct lruvec *lruvec;
|
|
|
|
int err = -EINVAL;
|
|
|
|
struct mem_cgroup *memcg = NULL;
|
|
|
|
|
|
|
|
if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
if (!mem_cgroup_disabled()) {
|
|
|
|
rcu_read_lock();
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
memcg = mem_cgroup_from_id(memcg_id);
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
if (!mem_cgroup_tryget(memcg))
|
2022-09-18 08:00:09 +00:00
|
|
|
memcg = NULL;
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
rcu_read_unlock();
|
|
|
|
|
|
|
|
if (!memcg)
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (memcg_id != mem_cgroup_id(memcg))
|
|
|
|
goto done;
|
|
|
|
|
|
|
|
lruvec = get_lruvec(memcg, nid);
|
|
|
|
|
|
|
|
if (swappiness < 0)
|
|
|
|
swappiness = get_swappiness(lruvec, sc);
|
|
|
|
else if (swappiness > 200)
|
|
|
|
goto done;
|
|
|
|
|
|
|
|
switch (cmd) {
|
|
|
|
case '+':
|
|
|
|
err = run_aging(lruvec, seq, sc, swappiness, opt);
|
|
|
|
break;
|
|
|
|
case '-':
|
|
|
|
err = run_eviction(lruvec, seq, sc, swappiness, opt);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
done:
|
|
|
|
mem_cgroup_put(memcg);
|
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2022-09-18 08:00:10 +00:00
|
|
|
/* see Documentation/admin-guide/mm/multigen_lru.rst for details */
|
2022-09-18 08:00:09 +00:00
|
|
|
static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
|
|
|
|
size_t len, loff_t *pos)
|
|
|
|
{
|
|
|
|
void *buf;
|
|
|
|
char *cur, *next;
|
|
|
|
unsigned int flags;
|
|
|
|
struct blk_plug plug;
|
|
|
|
int err = -EINVAL;
|
|
|
|
struct scan_control sc = {
|
|
|
|
.may_writepage = true,
|
|
|
|
.may_unmap = true,
|
|
|
|
.may_swap = true,
|
|
|
|
.reclaim_idx = MAX_NR_ZONES - 1,
|
|
|
|
.gfp_mask = GFP_KERNEL,
|
|
|
|
};
|
|
|
|
|
|
|
|
buf = kvmalloc(len + 1, GFP_KERNEL);
|
|
|
|
if (!buf)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
if (copy_from_user(buf, src, len)) {
|
|
|
|
kvfree(buf);
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
|
|
|
|
set_task_reclaim_state(current, &sc.reclaim_state);
|
|
|
|
flags = memalloc_noreclaim_save();
|
|
|
|
blk_start_plug(&plug);
|
mm: multi-gen LRU: clarify scan_control flags
Among the flags in scan_control:
1. sc->may_swap, which indicates swap constraint due to memsw.max, is
supported as usual.
2. sc->proactive, which indicates reclaim by memory.reclaim, may not
opportunistically skip the aging path, since it is considered less
latency sensitive.
3. !(sc->gfp_mask & __GFP_IO), which indicates IO constraint, lowers
swappiness to prioritize file LRU, since clean file folios are more
likely to exist.
4. sc->may_writepage and sc->may_unmap, which indicates opportunistic
reclaim, are rejected, since unmapped clean folios are already
prioritized. Scanning for more of them is likely futile and can
cause high reclaim latency when there is a large number of memcgs.
The rest are handled by the existing code.
Link: https://lkml.kernel.org/r/20221222041905.2431096-8-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:05 +00:00
|
|
|
if (!set_mm_walk(NULL, true)) {
|
2022-09-18 08:00:09 +00:00
|
|
|
err = -ENOMEM;
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
|
|
|
next = buf;
|
|
|
|
next[len] = '\0';
|
|
|
|
|
|
|
|
while ((cur = strsep(&next, ",;\n"))) {
|
|
|
|
int n;
|
|
|
|
int end;
|
|
|
|
char cmd;
|
|
|
|
unsigned int memcg_id;
|
|
|
|
unsigned int nid;
|
|
|
|
unsigned long seq;
|
|
|
|
unsigned int swappiness = -1;
|
|
|
|
unsigned long opt = -1;
|
|
|
|
|
|
|
|
cur = skip_spaces(cur);
|
|
|
|
if (!*cur)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
|
|
|
|
&seq, &end, &swappiness, &end, &opt, &end);
|
|
|
|
if (n < 4 || cur[end]) {
|
|
|
|
err = -EINVAL;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
|
|
|
|
if (err)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
done:
|
|
|
|
clear_mm_walk();
|
|
|
|
blk_finish_plug(&plug);
|
|
|
|
memalloc_noreclaim_restore(flags);
|
|
|
|
set_task_reclaim_state(current, NULL);
|
|
|
|
|
|
|
|
kvfree(buf);
|
|
|
|
|
|
|
|
return err ? : len;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int lru_gen_seq_open(struct inode *inode, struct file *file)
|
|
|
|
{
|
|
|
|
return seq_open(file, &lru_gen_seq_ops);
|
|
|
|
}
|
|
|
|
|
|
|
|
static const struct file_operations lru_gen_rw_fops = {
|
|
|
|
.open = lru_gen_seq_open,
|
|
|
|
.read = seq_read,
|
|
|
|
.write = lru_gen_seq_write,
|
|
|
|
.llseek = seq_lseek,
|
|
|
|
.release = seq_release,
|
|
|
|
};
|
|
|
|
|
|
|
|
static const struct file_operations lru_gen_ro_fops = {
|
|
|
|
.open = lru_gen_seq_open,
|
|
|
|
.read = seq_read,
|
|
|
|
.llseek = seq_lseek,
|
|
|
|
.release = seq_release,
|
|
|
|
};
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
/******************************************************************************
|
|
|
|
* initialization
|
|
|
|
******************************************************************************/
|
|
|
|
|
|
|
|
void lru_gen_init_lruvec(struct lruvec *lruvec)
|
|
|
|
{
|
2022-09-18 08:00:08 +00:00
|
|
|
int i;
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
int gen, type, zone;
|
mm: multi-gen LRU: rename lru_gen_struct to lru_gen_folio
Patch series "mm: multi-gen LRU: memcg LRU", v3.
Overview
========
An memcg LRU is a per-node LRU of memcgs. It is also an LRU of LRUs,
since each node and memcg combination has an LRU of folios (see
mem_cgroup_lruvec()).
Its goal is to improve the scalability of global reclaim, which is
critical to system-wide memory overcommit in data centers. Note that
memcg reclaim is currently out of scope.
Its memory bloat is a pointer to each lruvec and negligible to each
pglist_data. In terms of traversing memcgs during global reclaim, it
improves the best-case complexity from O(n) to O(1) and does not affect
the worst-case complexity O(n). Therefore, on average, it has a sublinear
complexity in contrast to the current linear complexity.
The basic structure of an memcg LRU can be understood by an analogy to
the active/inactive LRU (of folios):
1. It has the young and the old (generations), i.e., the counterparts
to the active and the inactive;
2. The increment of max_seq triggers promotion, i.e., the counterpart
to activation;
3. Other events trigger similar operations, e.g., offlining an memcg
triggers demotion, i.e., the counterpart to deactivation.
In terms of global reclaim, it has two distinct features:
1. Sharding, which allows each thread to start at a random memcg (in
the old generation) and improves parallelism;
2. Eventual fairness, which allows direct reclaim to bail out at will
and reduces latency without affecting fairness over some time.
The commit message in patch 6 details the workflow:
https://lore.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com/
The following is a simple test to quickly verify its effectiveness.
Test design:
1. Create multiple memcgs.
2. Each memcg contains a job (fio).
3. All jobs access the same amount of memory randomly.
4. The system does not experience global memory pressure.
5. Periodically write to the root memory.reclaim.
Desired outcome:
1. All memcgs have similar pgsteal counts, i.e., stddev(pgsteal)
over mean(pgsteal) is close to 0%.
2. The total pgsteal is close to the total requested through
memory.reclaim, i.e., sum(pgsteal) over sum(requested) is close
to 100%.
Actual outcome [1]:
MGLRU off MGLRU on
stddev(pgsteal) / mean(pgsteal) 75% 20%
sum(pgsteal) / sum(requested) 425% 95%
####################################################################
MEMCGS=128
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
mkdir /sys/fs/cgroup/memcg$memcg
done
start() {
echo $BASHPID > /sys/fs/cgroup/memcg$memcg/cgroup.procs
fio -name=memcg$memcg --numjobs=1 --ioengine=mmap \
--filename=/dev/zero --size=1920M --rw=randrw \
--rate=64m,64m --random_distribution=random \
--fadvise_hint=0 --time_based --runtime=10h \
--group_reporting --minimal
}
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
start &
done
sleep 600
for ((i = 0; i < 600; i++)); do
echo 256m >/sys/fs/cgroup/memory.reclaim
sleep 6
done
for ((memcg = 0; memcg < $MEMCGS; memcg++)); do
grep "pgsteal " /sys/fs/cgroup/memcg$memcg/memory.stat
done
####################################################################
[1]: This was obtained from running the above script (touches less
than 256GB memory) on an EPYC 7B13 with 512GB DRAM for over an
hour.
This patch (of 8):
The new name lru_gen_folio will be more distinct from the coming
lru_gen_memcg.
Link: https://lkml.kernel.org/r/20221222041905.2431096-1-yuzhao@google.com
Link: https://lkml.kernel.org/r/20221222041905.2431096-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:18:59 +00:00
|
|
|
struct lru_gen_folio *lrugen = &lruvec->lrugen;
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
|
|
|
|
lrugen->max_seq = MIN_NR_GENS + 1;
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
lrugen->enabled = lru_gen_enabled();
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
|
2022-09-18 08:00:08 +00:00
|
|
|
for (i = 0; i <= MIN_NR_GENS + 1; i++)
|
|
|
|
lrugen->timestamps[i] = jiffies;
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
for_each_gen_type_zone(gen, type, zone)
|
2022-12-22 04:19:00 +00:00
|
|
|
INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]);
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
|
|
|
lruvec->mm_state.seq = MIN_NR_GENS;
|
|
|
|
init_waitqueue_head(&lruvec->mm_state.wait);
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef CONFIG_MEMCG
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
|
|
|
|
void lru_gen_init_pgdat(struct pglist_data *pgdat)
|
|
|
|
{
|
|
|
|
int i, j;
|
|
|
|
|
|
|
|
spin_lock_init(&pgdat->memcg_lru.lock);
|
|
|
|
|
|
|
|
for (i = 0; i < MEMCG_NR_GENS; i++) {
|
|
|
|
for (j = 0; j < MEMCG_NR_BINS; j++)
|
|
|
|
INIT_HLIST_NULLS_HEAD(&pgdat->memcg_lru.fifo[i][j], i);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
void lru_gen_init_memcg(struct mem_cgroup *memcg)
|
|
|
|
{
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
INIT_LIST_HEAD(&memcg->mm_list.fifo);
|
|
|
|
spin_lock_init(&memcg->mm_list.lock);
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void lru_gen_exit_memcg(struct mem_cgroup *memcg)
|
|
|
|
{
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
int i;
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
int nid;
|
|
|
|
|
2023-01-18 00:18:25 +00:00
|
|
|
VM_WARN_ON_ONCE(!list_empty(&memcg->mm_list.fifo));
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
for_each_node(nid) {
|
|
|
|
struct lruvec *lruvec = get_lruvec(memcg, nid);
|
|
|
|
|
2023-01-18 00:18:25 +00:00
|
|
|
VM_WARN_ON_ONCE(lruvec->mm_state.nr_walkers);
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
|
|
|
|
sizeof(lruvec->lrugen.nr_pages)));
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
|
2023-01-18 00:18:25 +00:00
|
|
|
lruvec->lrugen.list.next = LIST_POISON1;
|
|
|
|
|
mm: multi-gen LRU: support page table walks
To further exploit spatial locality, the aging prefers to walk page tables
to search for young PTEs and promote hot pages. A kill switch will be
added in the next patch to disable this behavior. When disabled, the
aging relies on the rmap only.
NB: this behavior has nothing similar with the page table scanning in the
2.4 kernel [1], which searches page tables for old PTEs, adds cold pages
to swapcache and unmaps them.
To avoid confusion, the term "iteration" specifically means the traversal
of an entire mm_struct list; the term "walk" will be applied to page
tables and the rmap, as usual.
An mm_struct list is maintained for each memcg, and an mm_struct follows
its owner task to the new memcg when this task is migrated. Given an
lruvec, the aging iterates lruvec_memcg()->mm_list and calls
walk_page_range() with each mm_struct on this list to promote hot pages
before it increments max_seq.
When multiple page table walkers iterate the same list, each of them gets
a unique mm_struct; therefore they can run concurrently. Page table
walkers ignore any misplaced pages, e.g., if an mm_struct was migrated,
pages it left in the previous memcg will not be promoted when its current
memcg is under reclaim. Similarly, page table walkers will not promote
pages from nodes other than the one under reclaim.
This patch uses the following optimizations when walking page tables:
1. It tracks the usage of mm_struct's between context switches so that
page table walkers can skip processes that have been sleeping since
the last iteration.
2. It uses generational Bloom filters to record populated branches so
that page table walkers can reduce their search space based on the
query results, e.g., to skip page tables containing mostly holes or
misplaced pages.
3. It takes advantage of the accessed bit in non-leaf PMD entries when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y.
4. It does not zigzag between a PGD table and the same PMD table
spanning multiple VMAs. IOW, it finishes all the VMAs within the
range of the same PMD table before it returns to a PGD table. This
improves the cache performance for workloads that have large
numbers of tiny VMAs [2], especially when CONFIG_PGTABLE_LEVELS=5.
Server benchmark results:
Single workload:
fio (buffered I/O): no change
Single workload:
memcached (anon): +[8, 10]%
Ops/sec KB/sec
patch1-7: 1147696.57 44640.29
patch1-8: 1245274.91 48435.66
Configurations:
no change
Client benchmark results:
kswapd profiles:
patch1-7
48.16% lzo1x_1_do_compress (real work)
8.20% page_vma_mapped_walk (overhead)
7.06% _raw_spin_unlock_irq
2.92% ptep_clear_flush
2.53% __zram_bvec_write
2.11% do_raw_spin_lock
2.02% memmove
1.93% lru_gen_look_around
1.56% free_unref_page_list
1.40% memset
patch1-8
49.44% lzo1x_1_do_compress (real work)
6.19% page_vma_mapped_walk (overhead)
5.97% _raw_spin_unlock_irq
3.13% get_pfn_folio
2.85% ptep_clear_flush
2.42% __zram_bvec_write
2.08% do_raw_spin_lock
1.92% memmove
1.44% alloc_zspage
1.36% memset
Configurations:
no change
Thanks to the following developers for their efforts [3].
kernel test robot <lkp@intel.com>
[1] https://lwn.net/Articles/23732/
[2] https://llvm.org/docs/ScudoHardenedAllocator.html
[3] https://lore.kernel.org/r/202204160827.ekEARWQo-lkp@intel.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-9-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:05 +00:00
|
|
|
for (i = 0; i < NR_BLOOM_FILTERS; i++) {
|
|
|
|
bitmap_free(lruvec->mm_state.filters[i]);
|
|
|
|
lruvec->mm_state.filters[i] = NULL;
|
|
|
|
}
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
}
|
|
|
|
}
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
|
|
|
|
#endif /* CONFIG_MEMCG */
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
|
|
|
|
static int __init init_lru_gen(void)
|
|
|
|
{
|
|
|
|
BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
|
|
|
|
BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
|
|
|
|
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:07 +00:00
|
|
|
if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
|
|
|
|
pr_err("lru_gen: failed to create sysfs group\n");
|
|
|
|
|
2022-09-18 08:00:09 +00:00
|
|
|
debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
|
|
|
|
debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
return 0;
|
|
|
|
};
|
|
|
|
late_initcall(init_lru_gen);
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
#else /* !CONFIG_LRU_GEN */
|
|
|
|
|
|
|
|
static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
}
|
|
|
|
|
|
|
|
static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *sc)
|
|
|
|
{
|
|
|
|
}
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:02 +00:00
|
|
|
#endif /* CONFIG_LRU_GEN */
|
|
|
|
|
2019-12-01 01:55:46 +00:00
|
|
|
static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
|
2013-02-23 00:32:19 +00:00
|
|
|
{
|
|
|
|
unsigned long nr[NR_LRU_LISTS];
|
2013-07-03 22:01:44 +00:00
|
|
|
unsigned long targets[NR_LRU_LISTS];
|
2013-02-23 00:32:19 +00:00
|
|
|
unsigned long nr_to_scan;
|
|
|
|
enum lru_list lru;
|
|
|
|
unsigned long nr_reclaimed = 0;
|
|
|
|
unsigned long nr_to_reclaim = sc->nr_to_reclaim;
|
mm: vmscan: fix extreme overreclaim and swap floods
During proactive reclaim, we sometimes observe severe overreclaim, with
several thousand times more pages reclaimed than requested.
This trace was obtained from shrink_lruvec() during such an instance:
prio:0 anon_cost:1141521 file_cost:7767
nr_reclaimed:4387406 nr_to_reclaim:1047 (or_factor:4190)
nr=[7161123 345 578 1111]
While he reclaimer requested 4M, vmscan reclaimed close to 16G, most of it
by swapping. These requests take over a minute, during which the write()
to memory.reclaim is unkillably stuck inside the kernel.
Digging into the source, this is caused by the proportional reclaim
bailout logic. This code tries to resolve a fundamental conflict: to
reclaim roughly what was requested, while also aging all LRUs fairly and
in accordance to their size, swappiness, refault rates etc. The way it
attempts fairness is that once the reclaim goal has been reached, it stops
scanning the LRUs with the smaller remaining scan targets, and adjusts the
remainder of the bigger LRUs according to how much of the smaller LRUs was
scanned. It then finishes scanning that remainder regardless of the
reclaim goal.
This works fine if priority levels are low and the LRU lists are
comparable in size. However, in this instance, the cgroup that is
targeted by proactive reclaim has almost no files left - they've already
been squeezed out by proactive reclaim earlier - and the remaining anon
pages are hot. Anon rotations cause the priority level to drop to 0,
which results in reclaim targeting all of anon (a lot) and all of file
(almost nothing). By the time reclaim decides to bail, it has scanned
most or all of the file target, and therefor must also scan most or all of
the enormous anon target. This target is thousands of times larger than
the reclaim goal, thus causing the overreclaim.
The bailout code hasn't changed in years, why is this failing now? The
most likely explanations are two other recent changes in anon reclaim:
1. Before the series starting with commit 5df741963d52 ("mm: fix LRU
balancing effect of new transparent huge pages"), the VM was
overall relatively reluctant to swap at all, even if swap was
configured. This means the LRU balancing code didn't come into play
as often as it does now, and mostly in high pressure situations
where pronounced swap activity wouldn't be as surprising.
2. For historic reasons, shrink_lruvec() loops on the scan targets of
all LRU lists except the active anon one, meaning it would bail if
the only remaining pages to scan were active anon - even if there
were a lot of them.
Before the series starting with commit ccc5dc67340c ("mm/vmscan:
make active/inactive ratio as 1:1 for anon lru"), most anon pages
would live on the active LRU; the inactive one would contain only a
handful of preselected reclaim candidates. After the series, anon
gets aged similarly to file, and the inactive list is the default
for new anon pages as well, making it often the much bigger list.
As a result, the VM is now more likely to actually finish large
anon targets than before.
Change the code such that only one SWAP_CLUSTER_MAX-sized nudge toward the
larger LRU lists is made before bailing out on a met reclaim goal.
This fixes the extreme overreclaim problem.
Fairness is more subtle and harder to evaluate. No obvious misbehavior
was observed on the test workload, in any case. Conceptually, fairness
should primarily be a cumulative effect from regular, lower priority
scans. Once the VM is in trouble and needs to escalate scan targets to
make forward progress, fairness needs to take a backseat. This is also
acknowledged by the myriad exceptions in get_scan_count(). This patch
makes fairness decrease gradually, as it keeps fairness work static over
increasing priority levels with growing scan targets. This should make
more sense - although we may have to re-visit the exact values.
Link: https://lkml.kernel.org/r/20220802162811.39216-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@surriel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-02 16:28:11 +00:00
|
|
|
bool proportional_reclaim;
|
2013-02-23 00:32:19 +00:00
|
|
|
struct blk_plug plug;
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
if (lru_gen_enabled() && !global_reclaim(sc)) {
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
lru_gen_shrink_lruvec(lruvec, sc);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2019-12-01 01:55:46 +00:00
|
|
|
get_scan_count(lruvec, sc, nr);
|
2013-02-23 00:32:19 +00:00
|
|
|
|
2013-07-03 22:01:44 +00:00
|
|
|
/* Record the original scan target for proportional adjustments later */
|
|
|
|
memcpy(targets, nr, sizeof(nr));
|
|
|
|
|
2014-06-04 23:10:49 +00:00
|
|
|
/*
|
|
|
|
* Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
|
|
|
|
* event that can occur when there is little memory pressure e.g.
|
|
|
|
* multiple streaming readers/writers. Hence, we do not abort scanning
|
|
|
|
* when the requested number of pages are reclaimed when scanning at
|
|
|
|
* DEF_PRIORITY on the assumption that the fact we are direct
|
|
|
|
* reclaiming implies that kswapd is not keeping up and it is best to
|
|
|
|
* do a batch of work at once. For memcg reclaim one check is made to
|
|
|
|
* abort proportional reclaim if either the file or anon lru has already
|
|
|
|
* dropped to zero at the first pass.
|
|
|
|
*/
|
mm: vmscan: fix extreme overreclaim and swap floods
During proactive reclaim, we sometimes observe severe overreclaim, with
several thousand times more pages reclaimed than requested.
This trace was obtained from shrink_lruvec() during such an instance:
prio:0 anon_cost:1141521 file_cost:7767
nr_reclaimed:4387406 nr_to_reclaim:1047 (or_factor:4190)
nr=[7161123 345 578 1111]
While he reclaimer requested 4M, vmscan reclaimed close to 16G, most of it
by swapping. These requests take over a minute, during which the write()
to memory.reclaim is unkillably stuck inside the kernel.
Digging into the source, this is caused by the proportional reclaim
bailout logic. This code tries to resolve a fundamental conflict: to
reclaim roughly what was requested, while also aging all LRUs fairly and
in accordance to their size, swappiness, refault rates etc. The way it
attempts fairness is that once the reclaim goal has been reached, it stops
scanning the LRUs with the smaller remaining scan targets, and adjusts the
remainder of the bigger LRUs according to how much of the smaller LRUs was
scanned. It then finishes scanning that remainder regardless of the
reclaim goal.
This works fine if priority levels are low and the LRU lists are
comparable in size. However, in this instance, the cgroup that is
targeted by proactive reclaim has almost no files left - they've already
been squeezed out by proactive reclaim earlier - and the remaining anon
pages are hot. Anon rotations cause the priority level to drop to 0,
which results in reclaim targeting all of anon (a lot) and all of file
(almost nothing). By the time reclaim decides to bail, it has scanned
most or all of the file target, and therefor must also scan most or all of
the enormous anon target. This target is thousands of times larger than
the reclaim goal, thus causing the overreclaim.
The bailout code hasn't changed in years, why is this failing now? The
most likely explanations are two other recent changes in anon reclaim:
1. Before the series starting with commit 5df741963d52 ("mm: fix LRU
balancing effect of new transparent huge pages"), the VM was
overall relatively reluctant to swap at all, even if swap was
configured. This means the LRU balancing code didn't come into play
as often as it does now, and mostly in high pressure situations
where pronounced swap activity wouldn't be as surprising.
2. For historic reasons, shrink_lruvec() loops on the scan targets of
all LRU lists except the active anon one, meaning it would bail if
the only remaining pages to scan were active anon - even if there
were a lot of them.
Before the series starting with commit ccc5dc67340c ("mm/vmscan:
make active/inactive ratio as 1:1 for anon lru"), most anon pages
would live on the active LRU; the inactive one would contain only a
handful of preselected reclaim candidates. After the series, anon
gets aged similarly to file, and the inactive list is the default
for new anon pages as well, making it often the much bigger list.
As a result, the VM is now more likely to actually finish large
anon targets than before.
Change the code such that only one SWAP_CLUSTER_MAX-sized nudge toward the
larger LRU lists is made before bailing out on a met reclaim goal.
This fixes the extreme overreclaim problem.
Fairness is more subtle and harder to evaluate. No obvious misbehavior
was observed on the test workload, in any case. Conceptually, fairness
should primarily be a cumulative effect from regular, lower priority
scans. Once the VM is in trouble and needs to escalate scan targets to
make forward progress, fairness needs to take a backseat. This is also
acknowledged by the myriad exceptions in get_scan_count(). This patch
makes fairness decrease gradually, as it keeps fairness work static over
increasing priority levels with growing scan targets. This should make
more sense - although we may have to re-visit the exact values.
Link: https://lkml.kernel.org/r/20220802162811.39216-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@surriel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-02 16:28:11 +00:00
|
|
|
proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
|
|
|
|
sc->priority == DEF_PRIORITY);
|
2014-06-04 23:10:49 +00:00
|
|
|
|
2013-02-23 00:32:19 +00:00
|
|
|
blk_start_plug(&plug);
|
|
|
|
while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
|
|
|
|
nr[LRU_INACTIVE_FILE]) {
|
2013-07-03 22:01:44 +00:00
|
|
|
unsigned long nr_anon, nr_file, percentage;
|
|
|
|
unsigned long nr_scanned;
|
|
|
|
|
2013-02-23 00:32:19 +00:00
|
|
|
for_each_evictable_lru(lru) {
|
|
|
|
if (nr[lru]) {
|
|
|
|
nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
|
|
|
|
nr[lru] -= nr_to_scan;
|
|
|
|
|
|
|
|
nr_reclaimed += shrink_list(lru, nr_to_scan,
|
2019-04-19 00:50:34 +00:00
|
|
|
lruvec, sc);
|
2013-02-23 00:32:19 +00:00
|
|
|
}
|
|
|
|
}
|
2013-07-03 22:01:44 +00:00
|
|
|
|
2016-12-03 01:26:48 +00:00
|
|
|
cond_resched();
|
|
|
|
|
mm: vmscan: fix extreme overreclaim and swap floods
During proactive reclaim, we sometimes observe severe overreclaim, with
several thousand times more pages reclaimed than requested.
This trace was obtained from shrink_lruvec() during such an instance:
prio:0 anon_cost:1141521 file_cost:7767
nr_reclaimed:4387406 nr_to_reclaim:1047 (or_factor:4190)
nr=[7161123 345 578 1111]
While he reclaimer requested 4M, vmscan reclaimed close to 16G, most of it
by swapping. These requests take over a minute, during which the write()
to memory.reclaim is unkillably stuck inside the kernel.
Digging into the source, this is caused by the proportional reclaim
bailout logic. This code tries to resolve a fundamental conflict: to
reclaim roughly what was requested, while also aging all LRUs fairly and
in accordance to their size, swappiness, refault rates etc. The way it
attempts fairness is that once the reclaim goal has been reached, it stops
scanning the LRUs with the smaller remaining scan targets, and adjusts the
remainder of the bigger LRUs according to how much of the smaller LRUs was
scanned. It then finishes scanning that remainder regardless of the
reclaim goal.
This works fine if priority levels are low and the LRU lists are
comparable in size. However, in this instance, the cgroup that is
targeted by proactive reclaim has almost no files left - they've already
been squeezed out by proactive reclaim earlier - and the remaining anon
pages are hot. Anon rotations cause the priority level to drop to 0,
which results in reclaim targeting all of anon (a lot) and all of file
(almost nothing). By the time reclaim decides to bail, it has scanned
most or all of the file target, and therefor must also scan most or all of
the enormous anon target. This target is thousands of times larger than
the reclaim goal, thus causing the overreclaim.
The bailout code hasn't changed in years, why is this failing now? The
most likely explanations are two other recent changes in anon reclaim:
1. Before the series starting with commit 5df741963d52 ("mm: fix LRU
balancing effect of new transparent huge pages"), the VM was
overall relatively reluctant to swap at all, even if swap was
configured. This means the LRU balancing code didn't come into play
as often as it does now, and mostly in high pressure situations
where pronounced swap activity wouldn't be as surprising.
2. For historic reasons, shrink_lruvec() loops on the scan targets of
all LRU lists except the active anon one, meaning it would bail if
the only remaining pages to scan were active anon - even if there
were a lot of them.
Before the series starting with commit ccc5dc67340c ("mm/vmscan:
make active/inactive ratio as 1:1 for anon lru"), most anon pages
would live on the active LRU; the inactive one would contain only a
handful of preselected reclaim candidates. After the series, anon
gets aged similarly to file, and the inactive list is the default
for new anon pages as well, making it often the much bigger list.
As a result, the VM is now more likely to actually finish large
anon targets than before.
Change the code such that only one SWAP_CLUSTER_MAX-sized nudge toward the
larger LRU lists is made before bailing out on a met reclaim goal.
This fixes the extreme overreclaim problem.
Fairness is more subtle and harder to evaluate. No obvious misbehavior
was observed on the test workload, in any case. Conceptually, fairness
should primarily be a cumulative effect from regular, lower priority
scans. Once the VM is in trouble and needs to escalate scan targets to
make forward progress, fairness needs to take a backseat. This is also
acknowledged by the myriad exceptions in get_scan_count(). This patch
makes fairness decrease gradually, as it keeps fairness work static over
increasing priority levels with growing scan targets. This should make
more sense - although we may have to re-visit the exact values.
Link: https://lkml.kernel.org/r/20220802162811.39216-1-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@surriel.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-02 16:28:11 +00:00
|
|
|
if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
|
2013-07-03 22:01:44 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* For kswapd and memcg, reclaim at least the number of pages
|
2014-06-04 23:10:49 +00:00
|
|
|
* requested. Ensure that the anon and file LRUs are scanned
|
2013-07-03 22:01:44 +00:00
|
|
|
* proportionally what was requested by get_scan_count(). We
|
|
|
|
* stop reclaiming one LRU and reduce the amount scanning
|
|
|
|
* proportional to the original scan target.
|
|
|
|
*/
|
|
|
|
nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
|
|
|
|
nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
|
|
|
|
|
2014-06-04 23:10:49 +00:00
|
|
|
/*
|
|
|
|
* It's just vindictive to attack the larger once the smaller
|
|
|
|
* has gone to zero. And given the way we stop scanning the
|
|
|
|
* smaller below, this makes sure that we only make one nudge
|
|
|
|
* towards proportionality once we've got nr_to_reclaim.
|
|
|
|
*/
|
|
|
|
if (!nr_file || !nr_anon)
|
|
|
|
break;
|
|
|
|
|
2013-07-03 22:01:44 +00:00
|
|
|
if (nr_file > nr_anon) {
|
|
|
|
unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
|
|
|
|
targets[LRU_ACTIVE_ANON] + 1;
|
|
|
|
lru = LRU_BASE;
|
|
|
|
percentage = nr_anon * 100 / scan_target;
|
|
|
|
} else {
|
|
|
|
unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
|
|
|
|
targets[LRU_ACTIVE_FILE] + 1;
|
|
|
|
lru = LRU_FILE;
|
|
|
|
percentage = nr_file * 100 / scan_target;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Stop scanning the smaller of the LRU */
|
|
|
|
nr[lru] = 0;
|
|
|
|
nr[lru + LRU_ACTIVE] = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Recalculate the other LRU scan count based on its original
|
|
|
|
* scan target and the percentage scanning already complete
|
|
|
|
*/
|
|
|
|
lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
|
|
|
|
nr_scanned = targets[lru] - nr[lru];
|
|
|
|
nr[lru] = targets[lru] * (100 - percentage) / 100;
|
|
|
|
nr[lru] -= min(nr[lru], nr_scanned);
|
|
|
|
|
|
|
|
lru += LRU_ACTIVE;
|
|
|
|
nr_scanned = targets[lru] - nr[lru];
|
|
|
|
nr[lru] = targets[lru] * (100 - percentage) / 100;
|
|
|
|
nr[lru] -= min(nr[lru], nr_scanned);
|
2013-02-23 00:32:19 +00:00
|
|
|
}
|
|
|
|
blk_finish_plug(&plug);
|
|
|
|
sc->nr_reclaimed += nr_reclaimed;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Even if we did not try to evict anon pages at all, we want to
|
|
|
|
* rebalance the anon lru active/inactive ratio.
|
|
|
|
*/
|
2021-09-02 21:59:23 +00:00
|
|
|
if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
|
|
|
|
inactive_is_low(lruvec, LRU_INACTIVE_ANON))
|
2013-02-23 00:32:19 +00:00
|
|
|
shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
|
|
|
|
sc, LRU_ACTIVE_ANON);
|
|
|
|
}
|
|
|
|
|
2012-05-29 22:06:20 +00:00
|
|
|
/* Use reclaim/compaction for costly allocs or under memory pressure */
|
2012-05-29 22:06:57 +00:00
|
|
|
static bool in_reclaim_compaction(struct scan_control *sc)
|
2012-05-29 22:06:20 +00:00
|
|
|
{
|
2012-12-12 00:00:31 +00:00
|
|
|
if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
|
2012-05-29 22:06:20 +00:00
|
|
|
(sc->order > PAGE_ALLOC_COSTLY_ORDER ||
|
2012-05-29 22:06:57 +00:00
|
|
|
sc->priority < DEF_PRIORITY - 2))
|
2012-05-29 22:06:20 +00:00
|
|
|
return true;
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2011-01-13 23:45:56 +00:00
|
|
|
/*
|
2012-05-29 22:06:20 +00:00
|
|
|
* Reclaim/compaction is used for high-order allocation requests. It reclaims
|
|
|
|
* order-0 pages before compacting the zone. should_continue_reclaim() returns
|
|
|
|
* true if more pages should be reclaimed such that when the page allocator
|
2020-06-03 23:01:21 +00:00
|
|
|
* calls try_to_compact_pages() that it will have enough free pages to succeed.
|
2012-05-29 22:06:20 +00:00
|
|
|
* It will give up earlier than that if there is difficulty reclaiming pages.
|
2011-01-13 23:45:56 +00:00
|
|
|
*/
|
2016-07-28 22:46:02 +00:00
|
|
|
static inline bool should_continue_reclaim(struct pglist_data *pgdat,
|
2011-01-13 23:45:56 +00:00
|
|
|
unsigned long nr_reclaimed,
|
|
|
|
struct scan_control *sc)
|
|
|
|
{
|
|
|
|
unsigned long pages_for_compaction;
|
|
|
|
unsigned long inactive_lru_pages;
|
2016-07-28 22:46:02 +00:00
|
|
|
int z;
|
2011-01-13 23:45:56 +00:00
|
|
|
|
|
|
|
/* If not in reclaim/compaction mode, stop */
|
2012-05-29 22:06:57 +00:00
|
|
|
if (!in_reclaim_compaction(sc))
|
2011-01-13 23:45:56 +00:00
|
|
|
return false;
|
|
|
|
|
2019-09-23 22:37:29 +00:00
|
|
|
/*
|
|
|
|
* Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
|
|
|
|
* number of pages that were scanned. This will return to the caller
|
|
|
|
* with the risk reclaim/compaction and the resulting allocation attempt
|
|
|
|
* fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
|
|
|
|
* allocations through requiring that the full LRU list has been scanned
|
|
|
|
* first, by assuming that zero delta of sc->nr_scanned means full LRU
|
|
|
|
* scan, but that approximation was wrong, and there were corner cases
|
|
|
|
* where always a non-zero amount of pages were scanned.
|
|
|
|
*/
|
|
|
|
if (!nr_reclaimed)
|
|
|
|
return false;
|
2011-01-13 23:45:56 +00:00
|
|
|
|
|
|
|
/* If compaction would go ahead or the allocation would succeed, stop */
|
2016-07-28 22:46:02 +00:00
|
|
|
for (z = 0; z <= sc->reclaim_idx; z++) {
|
|
|
|
struct zone *zone = &pgdat->node_zones[z];
|
2016-09-01 23:14:55 +00:00
|
|
|
if (!managed_zone(zone))
|
2016-07-28 22:46:02 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
|
2016-10-07 23:57:41 +00:00
|
|
|
case COMPACT_SUCCESS:
|
2016-07-28 22:46:02 +00:00
|
|
|
case COMPACT_CONTINUE:
|
|
|
|
return false;
|
|
|
|
default:
|
|
|
|
/* check next zone */
|
|
|
|
;
|
|
|
|
}
|
2011-01-13 23:45:56 +00:00
|
|
|
}
|
2019-09-23 22:37:26 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* If we have not reclaimed enough pages for compaction and the
|
|
|
|
* inactive lists are large enough, continue reclaiming
|
|
|
|
*/
|
|
|
|
pages_for_compaction = compact_gap(sc->order);
|
|
|
|
inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
|
2021-09-02 21:59:26 +00:00
|
|
|
if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
|
2019-09-23 22:37:26 +00:00
|
|
|
inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
|
|
|
|
|
2019-09-23 22:37:29 +00:00
|
|
|
return inactive_lru_pages > pages_for_compaction;
|
2011-01-13 23:45:56 +00:00
|
|
|
}
|
|
|
|
|
2019-12-01 01:55:49 +00:00
|
|
|
static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2019-12-01 01:55:49 +00:00
|
|
|
struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
|
2019-12-01 01:55:43 +00:00
|
|
|
struct mem_cgroup *memcg;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2019-12-01 01:55:49 +00:00
|
|
|
memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
|
2019-12-01 01:55:43 +00:00
|
|
|
do {
|
2019-12-01 01:55:46 +00:00
|
|
|
struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
|
2019-12-01 01:55:43 +00:00
|
|
|
unsigned long reclaimed;
|
|
|
|
unsigned long scanned;
|
2012-01-13 01:17:59 +00:00
|
|
|
|
2020-09-04 23:35:27 +00:00
|
|
|
/*
|
|
|
|
* This loop can become CPU-bound when target memcgs
|
|
|
|
* aren't eligible for reclaim - either because they
|
|
|
|
* don't have any reclaimable pages, or because their
|
|
|
|
* memory is explicitly protected. Avoid soft lockups.
|
|
|
|
*/
|
|
|
|
cond_resched();
|
|
|
|
|
2020-08-07 06:22:05 +00:00
|
|
|
mem_cgroup_calculate_protection(target_memcg, memcg);
|
|
|
|
|
mm: memcg: fix stale protection of reclaim target memcg
Patch series "mm: memcg: fix protection of reclaim target memcg", v3.
This series fixes a bug in calculating the protection of the reclaim
target memcg where we end up using stale effective protection values from
the last reclaim operation, instead of completely ignoring the protection
of the reclaim target as intended. More detailed explanation and examples
in patch 1, which includes the fix. Patches 2 & 3 introduce a selftest
case that catches the bug.
This patch (of 3):
When we are doing memcg reclaim, the intended behavior is that we
ignore any protection (memory.min, memory.low) of the target memcg (but
not its children). Ever since the patch pointed to by the "Fixes" tag,
we actually read a stale value for the target memcg protection when
deciding whether to skip the memcg or not because it is protected. If
the stale value happens to be high enough, we don't reclaim from the
target memcg.
Essentially, in some cases we may falsely skip reclaiming from the
target memcg of reclaim because we read a stale protection value from
last time we reclaimed from it.
During reclaim, mem_cgroup_calculate_protection() is used to determine the
effective protection (emin and elow) values of a memcg. The protection of
the reclaim target is ignored, but we cannot set their effective
protection to 0 due to a limitation of the current implementation (see
comment in mem_cgroup_protection()). Instead, we leave their effective
protection values unchaged, and later ignore it in
mem_cgroup_protection().
However, mem_cgroup_protection() is called later in
shrink_lruvec()->get_scan_count(), which is after the
mem_cgroup_below_{min/low}() checks in shrink_node_memcgs(). As a result,
the stale effective protection values of the target memcg may lead us to
skip reclaiming from the target memcg entirely, before calling
shrink_lruvec(). This can be even worse with recursive protection, where
the stale target memcg protection can be higher than its standalone
protection. See two examples below (a similar version of example (a) is
added to test_memcontrol in a later patch).
(a) A simple example with proactive reclaim is as follows. Consider the
following hierarchy:
ROOT
|
A
|
B (memory.min = 10M)
Consider the following scenario:
- B has memory.current = 10M.
- The system undergoes global reclaim (or memcg reclaim in A).
- In shrink_node_memcgs():
- mem_cgroup_calculate_protection() calculates the effective min (emin)
of B as 10M.
- mem_cgroup_below_min() returns true for B, we do not reclaim from B.
- Now if we want to reclaim 5M from B using proactive reclaim
(memory.reclaim), we should be able to, as the protection of the
target memcg should be ignored.
- In shrink_node_memcgs():
- mem_cgroup_calculate_protection() immediately returns for B without
doing anything, as B is the target memcg, relying on
mem_cgroup_protection() to ignore B's stale effective min (still 10M).
- mem_cgroup_below_min() reads the stale effective min for B and we
skip it instead of ignoring its protection as intended, as we never
reach mem_cgroup_protection().
(b) An more complex example with recursive protection is as follows.
Consider the following hierarchy with memory_recursiveprot:
ROOT
|
A (memory.min = 50M)
|
B (memory.min = 10M, memory.high = 40M)
Consider the following scenario:
- B has memory.current = 35M.
- The system undergoes global reclaim (target memcg is NULL).
- B will have an effective min of 50M (all of A's unclaimed protection).
- B will not be reclaimed from.
- Now allocate 10M more memory in B, pushing it above it's high limit.
- The system undergoes memcg reclaim from B (target memcg is B).
- Like example (a), we do nothing in mem_cgroup_calculate_protection(),
then call mem_cgroup_below_min(), which will read the stale effective
min for B (50M) and skip it. In this case, it's even worse because we
are not just considering B's standalone protection (10M), but we are
reading a much higher stale protection (50M) which will cause us to not
reclaim from B at all.
This is an artifact of commit 45c7f7e1ef17 ("mm, memcg: decouple
e{low,min} state mutations from protection checks") which made
mem_cgroup_calculate_protection() only change the state without returning
any value. Before that commit, we used to return MEMCG_PROT_NONE for the
target memcg, which would cause us to skip the
mem_cgroup_below_{min/low}() checks. After that commit we do not return
anything and we end up checking the min & low effective protections for
the target memcg, which are stale.
Update mem_cgroup_supports_protection() to also check if we are reclaiming
from the target, and rename it to mem_cgroup_unprotected() (now returns
true if we should not protect the memcg, much simpler logic).
Link: https://lkml.kernel.org/r/20221202031512.1365483-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20221202031512.1365483-2-yosryahmed@google.com
Fixes: 45c7f7e1ef17 ("mm, memcg: decouple e{low,min} state mutations from protection checks")
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Chris Down <chris@chrisdown.name>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vasily Averin <vasily.averin@linux.dev>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-02 03:15:10 +00:00
|
|
|
if (mem_cgroup_below_min(target_memcg, memcg)) {
|
2019-12-01 01:55:43 +00:00
|
|
|
/*
|
|
|
|
* Hard protection.
|
|
|
|
* If there is no reclaimable memory, OOM.
|
|
|
|
*/
|
|
|
|
continue;
|
mm: memcg: fix stale protection of reclaim target memcg
Patch series "mm: memcg: fix protection of reclaim target memcg", v3.
This series fixes a bug in calculating the protection of the reclaim
target memcg where we end up using stale effective protection values from
the last reclaim operation, instead of completely ignoring the protection
of the reclaim target as intended. More detailed explanation and examples
in patch 1, which includes the fix. Patches 2 & 3 introduce a selftest
case that catches the bug.
This patch (of 3):
When we are doing memcg reclaim, the intended behavior is that we
ignore any protection (memory.min, memory.low) of the target memcg (but
not its children). Ever since the patch pointed to by the "Fixes" tag,
we actually read a stale value for the target memcg protection when
deciding whether to skip the memcg or not because it is protected. If
the stale value happens to be high enough, we don't reclaim from the
target memcg.
Essentially, in some cases we may falsely skip reclaiming from the
target memcg of reclaim because we read a stale protection value from
last time we reclaimed from it.
During reclaim, mem_cgroup_calculate_protection() is used to determine the
effective protection (emin and elow) values of a memcg. The protection of
the reclaim target is ignored, but we cannot set their effective
protection to 0 due to a limitation of the current implementation (see
comment in mem_cgroup_protection()). Instead, we leave their effective
protection values unchaged, and later ignore it in
mem_cgroup_protection().
However, mem_cgroup_protection() is called later in
shrink_lruvec()->get_scan_count(), which is after the
mem_cgroup_below_{min/low}() checks in shrink_node_memcgs(). As a result,
the stale effective protection values of the target memcg may lead us to
skip reclaiming from the target memcg entirely, before calling
shrink_lruvec(). This can be even worse with recursive protection, where
the stale target memcg protection can be higher than its standalone
protection. See two examples below (a similar version of example (a) is
added to test_memcontrol in a later patch).
(a) A simple example with proactive reclaim is as follows. Consider the
following hierarchy:
ROOT
|
A
|
B (memory.min = 10M)
Consider the following scenario:
- B has memory.current = 10M.
- The system undergoes global reclaim (or memcg reclaim in A).
- In shrink_node_memcgs():
- mem_cgroup_calculate_protection() calculates the effective min (emin)
of B as 10M.
- mem_cgroup_below_min() returns true for B, we do not reclaim from B.
- Now if we want to reclaim 5M from B using proactive reclaim
(memory.reclaim), we should be able to, as the protection of the
target memcg should be ignored.
- In shrink_node_memcgs():
- mem_cgroup_calculate_protection() immediately returns for B without
doing anything, as B is the target memcg, relying on
mem_cgroup_protection() to ignore B's stale effective min (still 10M).
- mem_cgroup_below_min() reads the stale effective min for B and we
skip it instead of ignoring its protection as intended, as we never
reach mem_cgroup_protection().
(b) An more complex example with recursive protection is as follows.
Consider the following hierarchy with memory_recursiveprot:
ROOT
|
A (memory.min = 50M)
|
B (memory.min = 10M, memory.high = 40M)
Consider the following scenario:
- B has memory.current = 35M.
- The system undergoes global reclaim (target memcg is NULL).
- B will have an effective min of 50M (all of A's unclaimed protection).
- B will not be reclaimed from.
- Now allocate 10M more memory in B, pushing it above it's high limit.
- The system undergoes memcg reclaim from B (target memcg is B).
- Like example (a), we do nothing in mem_cgroup_calculate_protection(),
then call mem_cgroup_below_min(), which will read the stale effective
min for B (50M) and skip it. In this case, it's even worse because we
are not just considering B's standalone protection (10M), but we are
reading a much higher stale protection (50M) which will cause us to not
reclaim from B at all.
This is an artifact of commit 45c7f7e1ef17 ("mm, memcg: decouple
e{low,min} state mutations from protection checks") which made
mem_cgroup_calculate_protection() only change the state without returning
any value. Before that commit, we used to return MEMCG_PROT_NONE for the
target memcg, which would cause us to skip the
mem_cgroup_below_{min/low}() checks. After that commit we do not return
anything and we end up checking the min & low effective protections for
the target memcg, which are stale.
Update mem_cgroup_supports_protection() to also check if we are reclaiming
from the target, and rename it to mem_cgroup_unprotected() (now returns
true if we should not protect the memcg, much simpler logic).
Link: https://lkml.kernel.org/r/20221202031512.1365483-1-yosryahmed@google.com
Link: https://lkml.kernel.org/r/20221202031512.1365483-2-yosryahmed@google.com
Fixes: 45c7f7e1ef17 ("mm, memcg: decouple e{low,min} state mutations from protection checks")
Signed-off-by: Yosry Ahmed <yosryahmed@google.com>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Chris Down <chris@chrisdown.name>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vasily Averin <vasily.averin@linux.dev>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-02 03:15:10 +00:00
|
|
|
} else if (mem_cgroup_below_low(target_memcg, memcg)) {
|
2019-12-01 01:55:43 +00:00
|
|
|
/*
|
|
|
|
* Soft protection.
|
|
|
|
* Respect the protection only as long as
|
|
|
|
* there is an unprotected supply
|
|
|
|
* of reclaimable memory from other cgroups.
|
|
|
|
*/
|
|
|
|
if (!sc->memcg_low_reclaim) {
|
|
|
|
sc->memcg_low_skipped = 1;
|
2018-06-08 00:07:46 +00:00
|
|
|
continue;
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 23:26:06 +00:00
|
|
|
}
|
2019-12-01 01:55:43 +00:00
|
|
|
memcg_memory_event(memcg, MEMCG_LOW);
|
|
|
|
}
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 23:26:06 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
reclaimed = sc->nr_reclaimed;
|
|
|
|
scanned = sc->nr_scanned;
|
2019-12-01 01:55:46 +00:00
|
|
|
|
|
|
|
shrink_lruvec(lruvec, sc);
|
memcg: add memory.pressure_level events
With this patch userland applications that want to maintain the
interactivity/memory allocation cost can use the pressure level
notifications. The levels are defined like this:
The "low" level means that the system is reclaiming memory for new
allocations. Monitoring this reclaiming activity might be useful for
maintaining cache level. Upon notification, the program (typically
"Activity Manager") might analyze vmstat and act in advance (i.e.
prematurely shutdown unimportant services).
The "medium" level means that the system is experiencing medium memory
pressure, the system might be making swap, paging out active file
caches, etc. Upon this event applications may decide to further analyze
vmstat/zoneinfo/memcg or internal memory usage statistics and free any
resources that can be easily reconstructed or re-read from a disk.
The "critical" level means that the system is actively thrashing, it is
about to out of memory (OOM) or even the in-kernel OOM killer is on its
way to trigger. Applications should do whatever they can to help the
system. It might be too late to consult with vmstat or any other
statistics, so it's advisable to take an immediate action.
The events are propagated upward until the event is handled, i.e. the
events are not pass-through. Here is what this means: for example you
have three cgroups: A->B->C. Now you set up an event listener on
cgroups A, B and C, and suppose group C experiences some pressure. In
this situation, only group C will receive the notification, i.e. groups
A and B will not receive it. This is done to avoid excessive
"broadcasting" of messages, which disturbs the system and which is
especially bad if we are low on memory or thrashing. So, organize the
cgroups wisely, or propagate the events manually (or, ask us to
implement the pass-through events, explaining why would you need them.)
Performance wise, the memory pressure notifications feature itself is
lightweight and does not require much of bookkeeping, in contrast to the
rest of memcg features. Unfortunately, as of current memcg
implementation, pages accounting is an inseparable part and cannot be
turned off. The good news is that there are some efforts[1] to improve
the situation; plus, implementing the same, fully API-compatible[2]
interface for CONFIG_MEMCG=n case (e.g. embedded) is also a viable
option, so it will not require any changes on the userland side.
[1] http://permalink.gmane.org/gmane.linux.kernel.cgroups/6291
[2] http://lkml.org/lkml/2013/2/21/454
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix CONFIG_CGROPUPS=n warnings]
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Leonid Moiseichuk <leonid.moiseichuk@nokia.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-29 22:08:31 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
|
|
|
|
sc->priority);
|
mm: vmscan: invoke slab shrinkers from shrink_zone()
The slab shrinkers are currently invoked from the zonelist walkers in
kswapd, direct reclaim, and zone reclaim, all of which roughly gauge the
eligible LRU pages and assemble a nodemask to pass to NUMA-aware
shrinkers, which then again have to walk over the nodemask. This is
redundant code, extra runtime work, and fairly inaccurate when it comes to
the estimation of actually scannable LRU pages. The code duplication will
only get worse when making the shrinkers cgroup-aware and requiring them
to have out-of-band cgroup hierarchy walks as well.
Instead, invoke the shrinkers from shrink_zone(), which is where all
reclaimers end up, to avoid this duplication.
Take the count for eligible LRU pages out of get_scan_count(), which
considers many more factors than just the availability of swap space, like
zone_reclaimable_pages() currently does. Accumulate the number over all
visited lruvecs to get the per-zone value.
Some nodes have multiple zones due to memory addressing restrictions. To
avoid putting too much pressure on the shrinkers, only invoke them once
for each such node, using the class zone of the allocation as the pivot
zone.
For now, this integrates the slab shrinking better into the reclaim logic
and gets rid of duplicative invocations from kswapd, direct reclaim, and
zone reclaim. It also prepares for cgroup-awareness, allowing
memcg-capable shrinkers to be added at the lruvec level without much
duplication of both code and runtime work.
This changes kswapd behavior, which used to invoke the shrinkers for each
zone, but with scan ratios gathered from the entire node, resulting in
meaningless pressure quantities on multi-zone nodes.
Zone reclaim behavior also changes. It used to shrink slabs until the
same amount of pages were shrunk as were reclaimed from the LRUs. Now it
merely invokes the shrinkers once with the zone's scan ratio, which makes
the shrinkers go easier on caches that implement aging and would prefer
feeding back pressure from recently used slab objects to unused LRU pages.
[vdavydov@parallels.com: assure class zone is populated]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 00:56:13 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
/* Record the group's reclaim efficiency */
|
2022-07-14 06:49:18 +00:00
|
|
|
if (!sc->proactive)
|
|
|
|
vmpressure(sc->gfp_mask, memcg, false,
|
|
|
|
sc->nr_scanned - scanned,
|
|
|
|
sc->nr_reclaimed - reclaimed);
|
memcg: add memory.pressure_level events
With this patch userland applications that want to maintain the
interactivity/memory allocation cost can use the pressure level
notifications. The levels are defined like this:
The "low" level means that the system is reclaiming memory for new
allocations. Monitoring this reclaiming activity might be useful for
maintaining cache level. Upon notification, the program (typically
"Activity Manager") might analyze vmstat and act in advance (i.e.
prematurely shutdown unimportant services).
The "medium" level means that the system is experiencing medium memory
pressure, the system might be making swap, paging out active file
caches, etc. Upon this event applications may decide to further analyze
vmstat/zoneinfo/memcg or internal memory usage statistics and free any
resources that can be easily reconstructed or re-read from a disk.
The "critical" level means that the system is actively thrashing, it is
about to out of memory (OOM) or even the in-kernel OOM killer is on its
way to trigger. Applications should do whatever they can to help the
system. It might be too late to consult with vmstat or any other
statistics, so it's advisable to take an immediate action.
The events are propagated upward until the event is handled, i.e. the
events are not pass-through. Here is what this means: for example you
have three cgroups: A->B->C. Now you set up an event listener on
cgroups A, B and C, and suppose group C experiences some pressure. In
this situation, only group C will receive the notification, i.e. groups
A and B will not receive it. This is done to avoid excessive
"broadcasting" of messages, which disturbs the system and which is
especially bad if we are low on memory or thrashing. So, organize the
cgroups wisely, or propagate the events manually (or, ask us to
implement the pass-through events, explaining why would you need them.)
Performance wise, the memory pressure notifications feature itself is
lightweight and does not require much of bookkeeping, in contrast to the
rest of memcg features. Unfortunately, as of current memcg
implementation, pages accounting is an inseparable part and cannot be
turned off. The good news is that there are some efforts[1] to improve
the situation; plus, implementing the same, fully API-compatible[2]
interface for CONFIG_MEMCG=n case (e.g. embedded) is also a viable
option, so it will not require any changes on the userland side.
[1] http://permalink.gmane.org/gmane.linux.kernel.cgroups/6291
[2] http://lkml.org/lkml/2013/2/21/454
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix CONFIG_CGROPUPS=n warnings]
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Glauber Costa <glommer@parallels.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Luiz Capitulino <lcapitulino@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Leonid Moiseichuk <leonid.moiseichuk@nokia.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Cc: John Stultz <john.stultz@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-29 22:08:31 +00:00
|
|
|
|
2019-12-01 01:55:49 +00:00
|
|
|
} while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
|
|
|
|
}
|
|
|
|
|
2020-01-31 06:14:08 +00:00
|
|
|
static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
|
2019-12-01 01:55:49 +00:00
|
|
|
{
|
|
|
|
struct reclaim_state *reclaim_state = current->reclaim_state;
|
|
|
|
unsigned long nr_reclaimed, nr_scanned;
|
2019-12-01 01:55:52 +00:00
|
|
|
struct lruvec *target_lruvec;
|
2019-12-01 01:55:49 +00:00
|
|
|
bool reclaimable = false;
|
|
|
|
|
mm: multi-gen LRU: per-node lru_gen_folio lists
For each node, memcgs are divided into two generations: the old and
the young. For each generation, memcgs are randomly sharded into
multiple bins to improve scalability. For each bin, an RCU hlist_nulls
is virtually divided into three segments: the head, the tail and the
default.
An onlining memcg is added to the tail of a random bin in the old
generation. The eviction starts at the head of a random bin in the old
generation. The per-node memcg generation counter, whose reminder (mod
2) indexes the old generation, is incremented when all its bins become
empty.
There are four operations:
1. MEMCG_LRU_HEAD, which moves an memcg to the head of a random bin in
its current generation (old or young) and updates its "seg" to
"head";
2. MEMCG_LRU_TAIL, which moves an memcg to the tail of a random bin in
its current generation (old or young) and updates its "seg" to
"tail";
3. MEMCG_LRU_OLD, which moves an memcg to the head of a random bin in
the old generation, updates its "gen" to "old" and resets its "seg"
to "default";
4. MEMCG_LRU_YOUNG, which moves an memcg to the tail of a random bin
in the young generation, updates its "gen" to "young" and resets
its "seg" to "default".
The events that trigger the above operations are:
1. Exceeding the soft limit, which triggers MEMCG_LRU_HEAD;
2. The first attempt to reclaim an memcg below low, which triggers
MEMCG_LRU_TAIL;
3. The first attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_TAIL;
4. The second attempt to reclaim an memcg below reclaimable size
threshold, which triggers MEMCG_LRU_YOUNG;
5. Attempting to reclaim an memcg below min, which triggers
MEMCG_LRU_YOUNG;
6. Finishing the aging on the eviction path, which triggers
MEMCG_LRU_YOUNG;
7. Offlining an memcg, which triggers MEMCG_LRU_OLD.
Note that memcg LRU only applies to global reclaim, and the
round-robin incrementing of their max_seq counters ensures the
eventual fairness to all eligible memcgs. For memcg reclaim, it still
relies on mem_cgroup_iter().
Link: https://lkml.kernel.org/r/20221222041905.2431096-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-22 04:19:04 +00:00
|
|
|
if (lru_gen_enabled() && global_reclaim(sc)) {
|
|
|
|
lru_gen_shrink_node(pgdat, sc);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2019-12-01 01:55:52 +00:00
|
|
|
target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
|
|
|
|
|
2019-12-01 01:55:49 +00:00
|
|
|
again:
|
|
|
|
memset(&sc->nr, 0, sizeof(sc->nr));
|
|
|
|
|
|
|
|
nr_reclaimed = sc->nr_reclaimed;
|
|
|
|
nr_scanned = sc->nr_scanned;
|
|
|
|
|
2022-09-18 08:00:00 +00:00
|
|
|
prepare_scan_count(pgdat, sc);
|
mm: vmscan: move file exhaustion detection to the node level
Patch series "mm: fix page aging across multiple cgroups".
When applications are put into unconfigured cgroups for memory accounting
purposes, the cgrouping itself should not change the behavior of the page
reclaim code. We expect the VM to reclaim the coldest pages in the
system. But right now the VM can reclaim hot pages in one cgroup while
there is eligible cold cache in others.
This is because one part of the reclaim algorithm isn't truly cgroup
hierarchy aware: the inactive/active list balancing. That is the part
that is supposed to protect hot cache data from one-off streaming IO.
The recursive cgroup reclaim scheme will scan and rotate the physical LRU
lists of each eligible cgroup at the same rate in a round-robin fashion,
thereby establishing a relative order among the pages of all those
cgroups. However, the inactive/active balancing decisions are made
locally within each cgroup, so when a cgroup is running low on cold pages,
its hot pages will get reclaimed - even when sibling cgroups have plenty
of cold cache eligible in the same reclaim run.
For example:
[root@ham ~]# head -n1 /proc/meminfo
MemTotal: 1016336 kB
[root@ham ~]# ./reclaimtest2.sh
Establishing 50M active files in cgroup A...
Hot pages cached: 12800/12800 workingset-a
Linearly scanning through 18G of file data in cgroup B:
real 0m4.269s
user 0m0.051s
sys 0m4.182s
Hot pages cached: 134/12800 workingset-a
The streaming IO in B, which doesn't benefit from caching at all, pushes
out most of the workingset in A.
Solution
This series fixes the problem by elevating inactive/active balancing
decisions to the toplevel of the reclaim run. This is either a cgroup
that hit its limit, or straight-up global reclaim if there is physical
memory pressure. From there, it takes a recursive view of the cgroup
subtree to decide whether page deactivation is necessary.
In the test above, the VM will then recognize that cgroup B has plenty of
eligible cold cache, and that the hot pages in A can be spared:
[root@ham ~]# ./reclaimtest2.sh
Establishing 50M active files in cgroup A...
Hot pages cached: 12800/12800 workingset-a
Linearly scanning through 18G of file data in cgroup B:
real 0m4.244s
user 0m0.064s
sys 0m4.177s
Hot pages cached: 12800/12800 workingset-a
Implementation
Whether active pages can be deactivated or not is influenced by two
factors: the inactive list dropping below a minimum size relative to the
active list, and the occurence of refaults.
This patch series first moves refault detection to the reclaim root, then
enforces the minimum inactive size based on a recursive view of the cgroup
tree's LRUs.
History
Note that this actually never worked correctly in Linux cgroups. In the
past it worked for global reclaim and leaf limit reclaim only (we used to
have two physical LRU linkages per page), but it never worked for
intermediate limit reclaim over multiple leaf cgroups.
We're noticing this now because 1) we're putting everything into cgroups
for accounting, not just the things we want to control and 2) we're moving
away from leaf limits that invoke reclaim on individual cgroups, toward
large tree reclaim, triggered by high-level limits, or physical memory
pressure that is influenced by local protections such as memory.low and
memory.min instead.
This patch (of 3):
When file pages are lower than the watermark on a node, we try to force
scan anonymous pages to counter-act the balancing algorithms preference
for new file pages when they are likely thrashing. This is a node-level
decision, but it's currently made each time we look at an lruvec. This is
unnecessarily expensive and also a layering violation that makes the code
harder to understand.
Clean this up by making the check once per node and setting a flag in the
scan_control.
Link: http://lkml.kernel.org/r/20191107205334.158354-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:56 +00:00
|
|
|
|
2019-12-01 01:55:49 +00:00
|
|
|
shrink_node_memcgs(pgdat, sc);
|
2014-08-06 23:06:15 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
if (reclaim_state) {
|
|
|
|
sc->nr_reclaimed += reclaim_state->reclaimed_slab;
|
|
|
|
reclaim_state->reclaimed_slab = 0;
|
|
|
|
}
|
2018-04-10 23:27:59 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
/* Record the subtree's reclaim efficiency */
|
2022-07-14 06:49:18 +00:00
|
|
|
if (!sc->proactive)
|
|
|
|
vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
|
|
|
|
sc->nr_scanned - nr_scanned,
|
|
|
|
sc->nr_reclaimed - nr_reclaimed);
|
2018-04-10 23:27:59 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
if (sc->nr_reclaimed - nr_reclaimed)
|
|
|
|
reclaimable = true;
|
2018-04-10 23:27:59 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
if (current_is_kswapd()) {
|
|
|
|
/*
|
|
|
|
* If reclaim is isolating dirty pages under writeback,
|
|
|
|
* it implies that the long-lived page allocation rate
|
|
|
|
* is exceeding the page laundering rate. Either the
|
|
|
|
* global limits are not being effective at throttling
|
|
|
|
* processes due to the page distribution throughout
|
|
|
|
* zones or there is heavy usage of a slow backing
|
|
|
|
* device. The only option is to throttle from reclaim
|
|
|
|
* context which is not ideal as there is no guarantee
|
|
|
|
* the dirtying process is throttled in the same way
|
|
|
|
* balance_dirty_pages() manages.
|
|
|
|
*
|
|
|
|
* Once a node is flagged PGDAT_WRITEBACK, kswapd will
|
|
|
|
* count the number of pages under pages flagged for
|
|
|
|
* immediate reclaim and stall if any are encountered
|
|
|
|
* in the nr_immediate check below.
|
|
|
|
*/
|
|
|
|
if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
|
|
|
|
set_bit(PGDAT_WRITEBACK, &pgdat->flags);
|
2018-04-10 23:27:59 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
/* Allow kswapd to start writing pages during reclaim.*/
|
|
|
|
if (sc->nr.unqueued_dirty == sc->nr.file_taken)
|
|
|
|
set_bit(PGDAT_DIRTY, &pgdat->flags);
|
mm/vmscan: don't mess with pgdat->flags in memcg reclaim
memcg reclaim may alter pgdat->flags based on the state of LRU lists in
cgroup and its children. PGDAT_WRITEBACK may force kswapd to sleep
congested_wait(), PGDAT_DIRTY may force kswapd to writeback filesystem
pages. But the worst here is PGDAT_CONGESTED, since it may force all
direct reclaims to stall in wait_iff_congested(). Note that only kswapd
have powers to clear any of these bits. This might just never happen if
cgroup limits configured that way. So all direct reclaims will stall as
long as we have some congested bdi in the system.
Leave all pgdat->flags manipulations to kswapd. kswapd scans the whole
pgdat, only kswapd can clear pgdat->flags once node is balanced, thus
it's reasonable to leave all decisions about node state to kswapd.
Why only kswapd? Why not allow to global direct reclaim change these
flags? It is because currently only kswapd can clear these flags. I'm
less worried about the case when PGDAT_CONGESTED falsely not set, and
more worried about the case when it falsely set. If direct reclaimer
sets PGDAT_CONGESTED, do we have guarantee that after the congestion
problem is sorted out, kswapd will be woken up and clear the flag? It
seems like there is no such guarantee. E.g. direct reclaimers may
eventually balance pgdat and kswapd simply won't wake up (see
wakeup_kswapd()).
Moving pgdat->flags manipulation to kswapd, means that cgroup2 recalim
now loses its congestion throttling mechanism. Add per-cgroup
congestion state and throttle cgroup2 reclaimers if memcg is in
congestion state.
Currently there is no need in per-cgroup PGDAT_WRITEBACK and PGDAT_DIRTY
bits since they alter only kswapd behavior.
The problem could be easily demonstrated by creating heavy congestion in
one cgroup:
echo "+memory" > /sys/fs/cgroup/cgroup.subtree_control
mkdir -p /sys/fs/cgroup/congester
echo 512M > /sys/fs/cgroup/congester/memory.max
echo $$ > /sys/fs/cgroup/congester/cgroup.procs
/* generate a lot of diry data on slow HDD */
while true; do dd if=/dev/zero of=/mnt/sdb/zeroes bs=1M count=1024; done &
....
while true; do dd if=/dev/zero of=/mnt/sdb/zeroes bs=1M count=1024; done &
and some job in another cgroup:
mkdir /sys/fs/cgroup/victim
echo 128M > /sys/fs/cgroup/victim/memory.max
# time cat /dev/sda > /dev/null
real 10m15.054s
user 0m0.487s
sys 1m8.505s
According to the tracepoint in wait_iff_congested(), the 'cat' spent 50%
of the time sleeping there.
With the patch, cat don't waste time anymore:
# time cat /dev/sda > /dev/null
real 5m32.911s
user 0m0.411s
sys 0m56.664s
[aryabinin@virtuozzo.com: congestion state should be per-node]
Link: http://lkml.kernel.org/r/20180406135215.10057-1-aryabinin@virtuozzo.com
[ayabinin@virtuozzo.com: make congestion state per-cgroup-per-node instead of just per-cgroup[
Link: http://lkml.kernel.org/r/20180406180254.8970-2-aryabinin@virtuozzo.com
Link: http://lkml.kernel.org/r/20180323152029.11084-5-aryabinin@virtuozzo.com
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-10 23:28:03 +00:00
|
|
|
|
2018-04-10 23:27:59 +00:00
|
|
|
/*
|
2020-08-12 01:33:26 +00:00
|
|
|
* If kswapd scans pages marked for immediate
|
2019-12-01 01:55:43 +00:00
|
|
|
* reclaim and under writeback (nr_immediate), it
|
|
|
|
* implies that pages are cycling through the LRU
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
* faster than they are written so forcibly stall
|
|
|
|
* until some pages complete writeback.
|
2018-04-10 23:27:59 +00:00
|
|
|
*/
|
2019-12-01 01:55:43 +00:00
|
|
|
if (sc->nr.immediate)
|
2021-11-05 20:42:42 +00:00
|
|
|
reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
|
2019-12-01 01:55:43 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
* Tag a node/memcg as congested if all the dirty pages were marked
|
|
|
|
* for writeback and immediate reclaim (counted in nr.congested).
|
2019-12-01 01:55:52 +00:00
|
|
|
*
|
2019-12-01 01:55:43 +00:00
|
|
|
* Legacy memcg will stall in page writeback so avoid forcibly
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
* stalling in reclaim_throttle().
|
2019-12-01 01:55:43 +00:00
|
|
|
*/
|
2019-12-01 01:55:52 +00:00
|
|
|
if ((current_is_kswapd() ||
|
|
|
|
(cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
|
2019-12-01 01:55:43 +00:00
|
|
|
sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
|
2019-12-01 01:55:52 +00:00
|
|
|
set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
|
2019-12-01 01:55:43 +00:00
|
|
|
|
|
|
|
/*
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
* Stall direct reclaim for IO completions if the lruvec is
|
|
|
|
* node is congested. Allow kswapd to continue until it
|
2019-12-01 01:55:43 +00:00
|
|
|
* starts encountering unqueued dirty pages or cycling through
|
|
|
|
* the LRU too quickly.
|
|
|
|
*/
|
2019-12-01 01:55:52 +00:00
|
|
|
if (!current_is_kswapd() && current_may_throttle() &&
|
|
|
|
!sc->hibernation_mode &&
|
|
|
|
test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
|
2018-04-10 23:27:59 +00:00
|
|
|
|
2019-12-01 01:55:43 +00:00
|
|
|
if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
|
|
|
|
sc))
|
|
|
|
goto again;
|
2014-08-06 23:06:15 +00:00
|
|
|
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
/*
|
|
|
|
* Kswapd gives up on balancing particular nodes after too
|
|
|
|
* many failures to reclaim anything from them and goes to
|
|
|
|
* sleep. On reclaim progress, reset the failure counter. A
|
|
|
|
* successful direct reclaim run will revive a dormant kswapd.
|
|
|
|
*/
|
|
|
|
if (reclaimable)
|
|
|
|
pgdat->kswapd_failures = 0;
|
2012-01-13 01:17:52 +00:00
|
|
|
}
|
|
|
|
|
mm, compaction: defer each zone individually instead of preferred zone
When direct sync compaction is often unsuccessful, it may become deferred
for some time to avoid further useless attempts, both sync and async.
Successful high-order allocations un-defer compaction, while further
unsuccessful compaction attempts prolong the compaction deferred period.
Currently the checking and setting deferred status is performed only on
the preferred zone of the allocation that invoked direct compaction. But
compaction itself is attempted on all eligible zones in the zonelist, so
the behavior is suboptimal and may lead both to scenarios where 1)
compaction is attempted uselessly, or 2) where it's not attempted despite
good chances of succeeding, as shown on the examples below:
1) A direct compaction with Normal preferred zone failed and set
deferred compaction for the Normal zone. Another unrelated direct
compaction with DMA32 as preferred zone will attempt to compact DMA32
zone even though the first compaction attempt also included DMA32 zone.
In another scenario, compaction with Normal preferred zone failed to
compact Normal zone, but succeeded in the DMA32 zone, so it will not
defer compaction. In the next attempt, it will try Normal zone which
will fail again, instead of skipping Normal zone and trying DMA32
directly.
2) Kswapd will balance DMA32 zone and reset defer status based on
watermarks looking good. A direct compaction with preferred Normal
zone will skip compaction of all zones including DMA32 because Normal
was still deferred. The allocation might have succeeded in DMA32, but
won't.
This patch makes compaction deferring work on individual zone basis
instead of preferred zone. For each zone, it checks compaction_deferred()
to decide if the zone should be skipped. If watermarks fail after
compacting the zone, defer_compaction() is called. The zone where
watermarks passed can still be deferred when the allocation attempt is
unsuccessful. When allocation is successful, compaction_defer_reset() is
called for the zone containing the allocated page. This approach should
approximate calling defer_compaction() only on zones where compaction was
attempted and did not yield allocated page. There might be corner cases
but that is inevitable as long as the decision to stop compacting dues not
guarantee that a page will be allocated.
Due to a new COMPACT_DEFERRED return value, some functions relying
implicitly on COMPACT_SKIPPED = 0 had to be updated, with comments made
more accurate. The did_some_progress output parameter of
__alloc_pages_direct_compact() is removed completely, as the caller
actually does not use it after compaction sets it - it is only considered
when direct reclaim sets it.
During testing on a two-node machine with a single very small Normal zone
on node 1, this patch has improved success rates in stress-highalloc
mmtests benchmark. The success here were previously made worse by commit
3a025760fc15 ("mm: page_alloc: spill to remote nodes before waking
kswapd") as kswapd was no longer resetting often enough the deferred
compaction for the Normal zone, and DMA32 zones on both nodes were thus
not considered for compaction. On different machine, success rates were
improved with __GFP_NO_KSWAPD allocations.
[akpm@linux-foundation.org: fix CONFIG_COMPACTION=n build]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-09 22:27:02 +00:00
|
|
|
/*
|
2016-10-07 23:58:03 +00:00
|
|
|
* Returns true if compaction should go ahead for a costly-order request, or
|
|
|
|
* the allocation would already succeed without compaction. Return false if we
|
|
|
|
* should reclaim first.
|
mm, compaction: defer each zone individually instead of preferred zone
When direct sync compaction is often unsuccessful, it may become deferred
for some time to avoid further useless attempts, both sync and async.
Successful high-order allocations un-defer compaction, while further
unsuccessful compaction attempts prolong the compaction deferred period.
Currently the checking and setting deferred status is performed only on
the preferred zone of the allocation that invoked direct compaction. But
compaction itself is attempted on all eligible zones in the zonelist, so
the behavior is suboptimal and may lead both to scenarios where 1)
compaction is attempted uselessly, or 2) where it's not attempted despite
good chances of succeeding, as shown on the examples below:
1) A direct compaction with Normal preferred zone failed and set
deferred compaction for the Normal zone. Another unrelated direct
compaction with DMA32 as preferred zone will attempt to compact DMA32
zone even though the first compaction attempt also included DMA32 zone.
In another scenario, compaction with Normal preferred zone failed to
compact Normal zone, but succeeded in the DMA32 zone, so it will not
defer compaction. In the next attempt, it will try Normal zone which
will fail again, instead of skipping Normal zone and trying DMA32
directly.
2) Kswapd will balance DMA32 zone and reset defer status based on
watermarks looking good. A direct compaction with preferred Normal
zone will skip compaction of all zones including DMA32 because Normal
was still deferred. The allocation might have succeeded in DMA32, but
won't.
This patch makes compaction deferring work on individual zone basis
instead of preferred zone. For each zone, it checks compaction_deferred()
to decide if the zone should be skipped. If watermarks fail after
compacting the zone, defer_compaction() is called. The zone where
watermarks passed can still be deferred when the allocation attempt is
unsuccessful. When allocation is successful, compaction_defer_reset() is
called for the zone containing the allocated page. This approach should
approximate calling defer_compaction() only on zones where compaction was
attempted and did not yield allocated page. There might be corner cases
but that is inevitable as long as the decision to stop compacting dues not
guarantee that a page will be allocated.
Due to a new COMPACT_DEFERRED return value, some functions relying
implicitly on COMPACT_SKIPPED = 0 had to be updated, with comments made
more accurate. The did_some_progress output parameter of
__alloc_pages_direct_compact() is removed completely, as the caller
actually does not use it after compaction sets it - it is only considered
when direct reclaim sets it.
During testing on a two-node machine with a single very small Normal zone
on node 1, this patch has improved success rates in stress-highalloc
mmtests benchmark. The success here were previously made worse by commit
3a025760fc15 ("mm: page_alloc: spill to remote nodes before waking
kswapd") as kswapd was no longer resetting often enough the deferred
compaction for the Normal zone, and DMA32 zones on both nodes were thus
not considered for compaction. On different machine, success rates were
improved with __GFP_NO_KSWAPD allocations.
[akpm@linux-foundation.org: fix CONFIG_COMPACTION=n build]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-09 22:27:02 +00:00
|
|
|
*/
|
2016-07-28 22:46:38 +00:00
|
|
|
static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
|
2012-01-13 01:19:45 +00:00
|
|
|
{
|
2016-07-28 22:45:46 +00:00
|
|
|
unsigned long watermark;
|
2016-10-07 23:58:03 +00:00
|
|
|
enum compact_result suitable;
|
2012-01-13 01:19:45 +00:00
|
|
|
|
2016-10-07 23:58:03 +00:00
|
|
|
suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
|
|
|
|
if (suitable == COMPACT_SUCCESS)
|
|
|
|
/* Allocation should succeed already. Don't reclaim. */
|
|
|
|
return true;
|
|
|
|
if (suitable == COMPACT_SKIPPED)
|
|
|
|
/* Compaction cannot yet proceed. Do reclaim. */
|
|
|
|
return false;
|
2012-01-13 01:19:45 +00:00
|
|
|
|
mm, compaction: defer each zone individually instead of preferred zone
When direct sync compaction is often unsuccessful, it may become deferred
for some time to avoid further useless attempts, both sync and async.
Successful high-order allocations un-defer compaction, while further
unsuccessful compaction attempts prolong the compaction deferred period.
Currently the checking and setting deferred status is performed only on
the preferred zone of the allocation that invoked direct compaction. But
compaction itself is attempted on all eligible zones in the zonelist, so
the behavior is suboptimal and may lead both to scenarios where 1)
compaction is attempted uselessly, or 2) where it's not attempted despite
good chances of succeeding, as shown on the examples below:
1) A direct compaction with Normal preferred zone failed and set
deferred compaction for the Normal zone. Another unrelated direct
compaction with DMA32 as preferred zone will attempt to compact DMA32
zone even though the first compaction attempt also included DMA32 zone.
In another scenario, compaction with Normal preferred zone failed to
compact Normal zone, but succeeded in the DMA32 zone, so it will not
defer compaction. In the next attempt, it will try Normal zone which
will fail again, instead of skipping Normal zone and trying DMA32
directly.
2) Kswapd will balance DMA32 zone and reset defer status based on
watermarks looking good. A direct compaction with preferred Normal
zone will skip compaction of all zones including DMA32 because Normal
was still deferred. The allocation might have succeeded in DMA32, but
won't.
This patch makes compaction deferring work on individual zone basis
instead of preferred zone. For each zone, it checks compaction_deferred()
to decide if the zone should be skipped. If watermarks fail after
compacting the zone, defer_compaction() is called. The zone where
watermarks passed can still be deferred when the allocation attempt is
unsuccessful. When allocation is successful, compaction_defer_reset() is
called for the zone containing the allocated page. This approach should
approximate calling defer_compaction() only on zones where compaction was
attempted and did not yield allocated page. There might be corner cases
but that is inevitable as long as the decision to stop compacting dues not
guarantee that a page will be allocated.
Due to a new COMPACT_DEFERRED return value, some functions relying
implicitly on COMPACT_SKIPPED = 0 had to be updated, with comments made
more accurate. The did_some_progress output parameter of
__alloc_pages_direct_compact() is removed completely, as the caller
actually does not use it after compaction sets it - it is only considered
when direct reclaim sets it.
During testing on a two-node machine with a single very small Normal zone
on node 1, this patch has improved success rates in stress-highalloc
mmtests benchmark. The success here were previously made worse by commit
3a025760fc15 ("mm: page_alloc: spill to remote nodes before waking
kswapd") as kswapd was no longer resetting often enough the deferred
compaction for the Normal zone, and DMA32 zones on both nodes were thus
not considered for compaction. On different machine, success rates were
improved with __GFP_NO_KSWAPD allocations.
[akpm@linux-foundation.org: fix CONFIG_COMPACTION=n build]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-09 22:27:02 +00:00
|
|
|
/*
|
2016-10-07 23:58:03 +00:00
|
|
|
* Compaction is already possible, but it takes time to run and there
|
|
|
|
* are potentially other callers using the pages just freed. So proceed
|
|
|
|
* with reclaim to make a buffer of free pages available to give
|
|
|
|
* compaction a reasonable chance of completing and allocating the page.
|
|
|
|
* Note that we won't actually reclaim the whole buffer in one attempt
|
|
|
|
* as the target watermark in should_continue_reclaim() is lower. But if
|
|
|
|
* we are already above the high+gap watermark, don't reclaim at all.
|
mm, compaction: defer each zone individually instead of preferred zone
When direct sync compaction is often unsuccessful, it may become deferred
for some time to avoid further useless attempts, both sync and async.
Successful high-order allocations un-defer compaction, while further
unsuccessful compaction attempts prolong the compaction deferred period.
Currently the checking and setting deferred status is performed only on
the preferred zone of the allocation that invoked direct compaction. But
compaction itself is attempted on all eligible zones in the zonelist, so
the behavior is suboptimal and may lead both to scenarios where 1)
compaction is attempted uselessly, or 2) where it's not attempted despite
good chances of succeeding, as shown on the examples below:
1) A direct compaction with Normal preferred zone failed and set
deferred compaction for the Normal zone. Another unrelated direct
compaction with DMA32 as preferred zone will attempt to compact DMA32
zone even though the first compaction attempt also included DMA32 zone.
In another scenario, compaction with Normal preferred zone failed to
compact Normal zone, but succeeded in the DMA32 zone, so it will not
defer compaction. In the next attempt, it will try Normal zone which
will fail again, instead of skipping Normal zone and trying DMA32
directly.
2) Kswapd will balance DMA32 zone and reset defer status based on
watermarks looking good. A direct compaction with preferred Normal
zone will skip compaction of all zones including DMA32 because Normal
was still deferred. The allocation might have succeeded in DMA32, but
won't.
This patch makes compaction deferring work on individual zone basis
instead of preferred zone. For each zone, it checks compaction_deferred()
to decide if the zone should be skipped. If watermarks fail after
compacting the zone, defer_compaction() is called. The zone where
watermarks passed can still be deferred when the allocation attempt is
unsuccessful. When allocation is successful, compaction_defer_reset() is
called for the zone containing the allocated page. This approach should
approximate calling defer_compaction() only on zones where compaction was
attempted and did not yield allocated page. There might be corner cases
but that is inevitable as long as the decision to stop compacting dues not
guarantee that a page will be allocated.
Due to a new COMPACT_DEFERRED return value, some functions relying
implicitly on COMPACT_SKIPPED = 0 had to be updated, with comments made
more accurate. The did_some_progress output parameter of
__alloc_pages_direct_compact() is removed completely, as the caller
actually does not use it after compaction sets it - it is only considered
when direct reclaim sets it.
During testing on a two-node machine with a single very small Normal zone
on node 1, this patch has improved success rates in stress-highalloc
mmtests benchmark. The success here were previously made worse by commit
3a025760fc15 ("mm: page_alloc: spill to remote nodes before waking
kswapd") as kswapd was no longer resetting often enough the deferred
compaction for the Normal zone, and DMA32 zones on both nodes were thus
not considered for compaction. On different machine, success rates were
improved with __GFP_NO_KSWAPD allocations.
[akpm@linux-foundation.org: fix CONFIG_COMPACTION=n build]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-10-09 22:27:02 +00:00
|
|
|
*/
|
2016-10-07 23:58:03 +00:00
|
|
|
watermark = high_wmark_pages(zone) + compact_gap(sc->order);
|
2012-01-13 01:19:45 +00:00
|
|
|
|
2016-10-07 23:58:03 +00:00
|
|
|
return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
|
2012-01-13 01:19:45 +00:00
|
|
|
}
|
|
|
|
|
2021-11-05 20:42:32 +00:00
|
|
|
static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
|
|
|
|
{
|
2021-11-05 20:42:49 +00:00
|
|
|
/*
|
|
|
|
* If reclaim is making progress greater than 12% efficiency then
|
|
|
|
* wake all the NOPROGRESS throttled tasks.
|
|
|
|
*/
|
|
|
|
if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
|
2021-11-05 20:42:32 +00:00
|
|
|
wait_queue_head_t *wqh;
|
|
|
|
|
|
|
|
wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
|
|
|
|
if (waitqueue_active(wqh))
|
|
|
|
wake_up(wqh);
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
* Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
|
|
|
|
* throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
|
|
|
|
* under writeback and marked for immediate reclaim at the tail of the
|
|
|
|
* LRU.
|
2021-11-05 20:42:32 +00:00
|
|
|
*/
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
if (current_is_kswapd() || cgroup_reclaim(sc))
|
2021-11-05 20:42:32 +00:00
|
|
|
return;
|
|
|
|
|
|
|
|
/* Throttle if making no progress at high prioities. */
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
if (sc->priority == 1 && !sc->nr_reclaimed)
|
2021-11-05 20:42:42 +00:00
|
|
|
reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
|
2021-11-05 20:42:32 +00:00
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* This is the direct reclaim path, for page-allocating processes. We only
|
|
|
|
* try to reclaim pages from zones which will satisfy the caller's allocation
|
|
|
|
* request.
|
|
|
|
*
|
|
|
|
* If a zone is deemed to be full of pinned pages then just give it a light
|
|
|
|
* scan then give up on it.
|
|
|
|
*/
|
2016-05-20 23:57:00 +00:00
|
|
|
static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2008-04-28 09:12:17 +00:00
|
|
|
struct zoneref *z;
|
2008-04-28 09:12:16 +00:00
|
|
|
struct zone *zone;
|
2013-09-24 22:27:41 +00:00
|
|
|
unsigned long nr_soft_reclaimed;
|
|
|
|
unsigned long nr_soft_scanned;
|
2014-04-07 22:36:59 +00:00
|
|
|
gfp_t orig_mask;
|
2016-07-28 22:45:53 +00:00
|
|
|
pg_data_t *last_pgdat = NULL;
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
pg_data_t *first_pgdat = NULL;
|
2008-02-07 08:14:37 +00:00
|
|
|
|
2012-03-21 23:34:00 +00:00
|
|
|
/*
|
|
|
|
* If the number of buffer_heads in the machine exceeds the maximum
|
|
|
|
* allowed level, force direct reclaim to scan the highmem zone as
|
|
|
|
* highmem pages could be pinning lowmem pages storing buffer_heads
|
|
|
|
*/
|
2014-04-07 22:36:59 +00:00
|
|
|
orig_mask = sc->gfp_mask;
|
2016-07-28 22:45:37 +00:00
|
|
|
if (buffer_heads_over_limit) {
|
2012-03-21 23:34:00 +00:00
|
|
|
sc->gfp_mask |= __GFP_HIGHMEM;
|
2016-07-28 22:46:38 +00:00
|
|
|
sc->reclaim_idx = gfp_zone(sc->gfp_mask);
|
2016-07-28 22:45:37 +00:00
|
|
|
}
|
2012-03-21 23:34:00 +00:00
|
|
|
|
2010-08-10 00:19:29 +00:00
|
|
|
for_each_zone_zonelist_nodemask(zone, z, zonelist,
|
2016-07-28 22:45:37 +00:00
|
|
|
sc->reclaim_idx, sc->nodemask) {
|
2008-02-07 08:14:37 +00:00
|
|
|
/*
|
|
|
|
* Take care memory controller reclaiming has small influence
|
|
|
|
* to global LRU.
|
|
|
|
*/
|
2019-12-01 01:55:40 +00:00
|
|
|
if (!cgroup_reclaim(sc)) {
|
2014-10-20 11:50:30 +00:00
|
|
|
if (!cpuset_zone_allowed(zone,
|
|
|
|
GFP_KERNEL | __GFP_HARDWALL))
|
2008-02-07 08:14:37 +00:00
|
|
|
continue;
|
2014-04-03 21:47:20 +00:00
|
|
|
|
2014-08-06 23:06:12 +00:00
|
|
|
/*
|
|
|
|
* If we already have plenty of memory free for
|
|
|
|
* compaction in this zone, don't free any more.
|
|
|
|
* Even though compaction is invoked for any
|
|
|
|
* non-zero order, only frequent costly order
|
|
|
|
* reclamation is disruptive enough to become a
|
|
|
|
* noticeable problem, like transparent huge
|
|
|
|
* page allocations.
|
|
|
|
*/
|
|
|
|
if (IS_ENABLED(CONFIG_COMPACTION) &&
|
|
|
|
sc->order > PAGE_ALLOC_COSTLY_ORDER &&
|
2016-07-28 22:46:38 +00:00
|
|
|
compaction_ready(zone, sc)) {
|
2014-08-06 23:06:12 +00:00
|
|
|
sc->compaction_ready = true;
|
|
|
|
continue;
|
2011-11-01 00:09:31 +00:00
|
|
|
}
|
2014-08-06 23:06:12 +00:00
|
|
|
|
2016-07-28 22:45:53 +00:00
|
|
|
/*
|
|
|
|
* Shrink each node in the zonelist once. If the
|
|
|
|
* zonelist is ordered by zone (not the default) then a
|
|
|
|
* node may be shrunk multiple times but in that case
|
|
|
|
* the user prefers lower zones being preserved.
|
|
|
|
*/
|
|
|
|
if (zone->zone_pgdat == last_pgdat)
|
|
|
|
continue;
|
|
|
|
|
2013-09-24 22:27:41 +00:00
|
|
|
/*
|
|
|
|
* This steals pages from memory cgroups over softlimit
|
|
|
|
* and returns the number of reclaimed pages and
|
|
|
|
* scanned pages. This works for global memory pressure
|
|
|
|
* and balancing, not for a memcg's limit.
|
|
|
|
*/
|
|
|
|
nr_soft_scanned = 0;
|
2016-07-28 22:46:05 +00:00
|
|
|
nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
|
2013-09-24 22:27:41 +00:00
|
|
|
sc->order, sc->gfp_mask,
|
|
|
|
&nr_soft_scanned);
|
|
|
|
sc->nr_reclaimed += nr_soft_reclaimed;
|
|
|
|
sc->nr_scanned += nr_soft_scanned;
|
2011-06-27 23:18:12 +00:00
|
|
|
/* need some check for avoid more shrink_zone() */
|
2008-02-07 08:14:37 +00:00
|
|
|
}
|
2006-09-26 06:31:27 +00:00
|
|
|
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
if (!first_pgdat)
|
|
|
|
first_pgdat = zone->zone_pgdat;
|
|
|
|
|
2016-07-28 22:45:53 +00:00
|
|
|
/* See comment about same check for global reclaim above */
|
|
|
|
if (zone->zone_pgdat == last_pgdat)
|
|
|
|
continue;
|
|
|
|
last_pgdat = zone->zone_pgdat;
|
2016-07-28 22:46:35 +00:00
|
|
|
shrink_node(zone->zone_pgdat, sc);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2011-11-01 00:09:33 +00:00
|
|
|
|
2021-12-31 21:10:09 +00:00
|
|
|
if (first_pgdat)
|
|
|
|
consider_reclaim_throttle(first_pgdat, sc);
|
mm: vmscan: Reduce throttling due to a failure to make progress
Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time. In
Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.
Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is
being made") introduced the problem although commit a19594ca4a8b
("mm/vmscan: increase the timeout if page reclaim is not making
progress") made it worse. Systems at or near an OOM state that cannot
be recovered must reach OOM quickly and memcg should kill tasks if a
memcg is near OOM.
To address this, only stall for the first zone in the zonelist, reduce
the timeout to 1 tick for VMSCAN_THROTTLE_NOPROGRESS and only stall if
the scan control nr_reclaimed is 0, kswapd is still active and there
were excessive pages pending for writeback. If kswapd has stopped
reclaiming due to excessive failures, do not stall at all so that OOM
triggers relatively quickly. Similarly, if an LRU is simply congested,
only lightly throttle similar to NOPROGRESS.
Alexey's original case was the most straight forward
for i in {1..3}; do tail /dev/zero; done
On vanilla 5.16-rc1, this test stalled heavily, after the patch the test
completes in a few seconds similar to 5.15.
Alexey's second test case added watching a youtube video while tail runs
10 times. On 5.15, playback only jitters slightly, 5.16-rc1 stalls a
lot with lots of frames missing and numerous audio glitches. With this
patch applies, the video plays similarly to 5.15.
[lkp@intel.com: Fix W=1 build warning]
Link: https://lore.kernel.org/r/99e779783d6c7fce96448a3402061b9dc1b3b602.camel@gmx.de
Link: https://lore.kernel.org/r/20211124011954.7cab9bb4@mail.inbox.lv
Link: https://lore.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lore.kernel.org/r/20211202150614.22440-1-mgorman@techsingularity.net
Link: https://linux-regtracking.leemhuis.info/regzbot/regression/20211124011954.7cab9bb4@mail.inbox.lv/
Reported-and-tested-by: Alexey Avramov <hakavlad@inbox.lv>
Reported-and-tested-by: Mike Galbraith <efault@gmx.de>
Reported-and-tested-by: Darrick J. Wong <djwong@kernel.org>
Reported-by: kernel test robot <lkp@intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Tracked-by: Thorsten Leemhuis <regressions@leemhuis.info>
Fixes: 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being made")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-12-02 15:06:14 +00:00
|
|
|
|
2014-04-07 22:36:59 +00:00
|
|
|
/*
|
|
|
|
* Restore to original mask to avoid the impact on the caller if we
|
|
|
|
* promoted it to __GFP_HIGHMEM.
|
|
|
|
*/
|
|
|
|
sc->gfp_mask = orig_mask;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2008-10-19 03:26:32 +00:00
|
|
|
|
mm: vmscan: detect file thrashing at the reclaim root
We use refault information to determine whether the cache workingset is
stable or transitioning, and dynamically adjust the inactive:active file
LRU ratio so as to maximize protection from one-off cache during stable
periods, and minimize IO during transitions.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, refaults only affect the
local LRU order in the cgroup in which they are occuring. As a result,
cache transitions can take longer in a cgrouped system as the active pages
of sibling cgroups aren't challenged when they should be.
[ Right now, this is somewhat theoretical, because the siblings, under
continued regular reclaim pressure, should eventually run out of
inactive pages - and since inactive:active *size* balancing is also
done on a cgroup-local level, we will challenge the active pages
eventually in most cases. But the next patch will move that relative
size enforcement to the reclaim root as well, and then this patch
here will be necessary to propagate refault pressure to siblings. ]
This patch moves refault detection to the root of reclaim. Instead of
remembering the cgroup owner of an evicted page, remember the cgroup that
caused the reclaim to happen. When refaults later occur, they'll
correctly influence the cross-cgroup LRU order that reclaim follows.
I.e. if global reclaim kicked out pages in some subgroup A/B/C, the
refault of those pages will challenge the global LRU order, and not just
the local order down inside C.
[hannes@cmpxchg.org: use page_memcg() instead of another lookup]
Link: http://lkml.kernel.org/r/20191115160722.GA309754@cmpxchg.org
Link: http://lkml.kernel.org/r/20191107205334.158354-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:59 +00:00
|
|
|
static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
|
2017-05-03 21:55:03 +00:00
|
|
|
{
|
mm: vmscan: detect file thrashing at the reclaim root
We use refault information to determine whether the cache workingset is
stable or transitioning, and dynamically adjust the inactive:active file
LRU ratio so as to maximize protection from one-off cache during stable
periods, and minimize IO during transitions.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, refaults only affect the
local LRU order in the cgroup in which they are occuring. As a result,
cache transitions can take longer in a cgrouped system as the active pages
of sibling cgroups aren't challenged when they should be.
[ Right now, this is somewhat theoretical, because the siblings, under
continued regular reclaim pressure, should eventually run out of
inactive pages - and since inactive:active *size* balancing is also
done on a cgroup-local level, we will challenge the active pages
eventually in most cases. But the next patch will move that relative
size enforcement to the reclaim root as well, and then this patch
here will be necessary to propagate refault pressure to siblings. ]
This patch moves refault detection to the root of reclaim. Instead of
remembering the cgroup owner of an evicted page, remember the cgroup that
caused the reclaim to happen. When refaults later occur, they'll
correctly influence the cross-cgroup LRU order that reclaim follows.
I.e. if global reclaim kicked out pages in some subgroup A/B/C, the
refault of those pages will challenge the global LRU order, and not just
the local order down inside C.
[hannes@cmpxchg.org: use page_memcg() instead of another lookup]
Link: http://lkml.kernel.org/r/20191115160722.GA309754@cmpxchg.org
Link: http://lkml.kernel.org/r/20191107205334.158354-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:59 +00:00
|
|
|
struct lruvec *target_lruvec;
|
|
|
|
unsigned long refaults;
|
2017-05-03 21:55:03 +00:00
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
if (lru_gen_enabled())
|
|
|
|
return;
|
|
|
|
|
mm: vmscan: detect file thrashing at the reclaim root
We use refault information to determine whether the cache workingset is
stable or transitioning, and dynamically adjust the inactive:active file
LRU ratio so as to maximize protection from one-off cache during stable
periods, and minimize IO during transitions.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, refaults only affect the
local LRU order in the cgroup in which they are occuring. As a result,
cache transitions can take longer in a cgrouped system as the active pages
of sibling cgroups aren't challenged when they should be.
[ Right now, this is somewhat theoretical, because the siblings, under
continued regular reclaim pressure, should eventually run out of
inactive pages - and since inactive:active *size* balancing is also
done on a cgroup-local level, we will challenge the active pages
eventually in most cases. But the next patch will move that relative
size enforcement to the reclaim root as well, and then this patch
here will be necessary to propagate refault pressure to siblings. ]
This patch moves refault detection to the root of reclaim. Instead of
remembering the cgroup owner of an evicted page, remember the cgroup that
caused the reclaim to happen. When refaults later occur, they'll
correctly influence the cross-cgroup LRU order that reclaim follows.
I.e. if global reclaim kicked out pages in some subgroup A/B/C, the
refault of those pages will challenge the global LRU order, and not just
the local order down inside C.
[hannes@cmpxchg.org: use page_memcg() instead of another lookup]
Link: http://lkml.kernel.org/r/20191115160722.GA309754@cmpxchg.org
Link: http://lkml.kernel.org/r/20191107205334.158354-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:55:59 +00:00
|
|
|
target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
|
2020-08-12 01:30:43 +00:00
|
|
|
refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
|
2022-08-08 00:56:45 +00:00
|
|
|
target_lruvec->refaults[WORKINGSET_ANON] = refaults;
|
2020-08-12 01:30:43 +00:00
|
|
|
refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
|
2022-08-08 00:56:45 +00:00
|
|
|
target_lruvec->refaults[WORKINGSET_FILE] = refaults;
|
2017-05-03 21:55:03 +00:00
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* This is the main entry point to direct page reclaim.
|
|
|
|
*
|
|
|
|
* If a full scan of the inactive list fails to free enough memory then we
|
|
|
|
* are "out of memory" and something needs to be killed.
|
|
|
|
*
|
|
|
|
* If the caller is !__GFP_FS then the probability of a failure is reasonably
|
|
|
|
* high - the zone may be full of dirty or under-writeback pages, which this
|
2009-09-23 17:37:09 +00:00
|
|
|
* caller can't do much about. We kick the writeback threads and take explicit
|
|
|
|
* naps in the hope that some of these pages can be written. But if the
|
|
|
|
* allocating task holds filesystem locks which prevent writeout this might not
|
|
|
|
* work, and the allocation attempt will fail.
|
page allocator: smarter retry of costly-order allocations
Because of page order checks in __alloc_pages(), hugepage (and similarly
large order) allocations will not retry unless explicitly marked
__GFP_REPEAT. However, the current retry logic is nearly an infinite
loop (or until reclaim does no progress whatsoever). For these costly
allocations, that seems like overkill and could potentially never
terminate. Mel observed that allowing current __GFP_REPEAT semantics for
hugepage allocations essentially killed the system. I believe this is
because we may continue to reclaim small orders of pages all over, but
never have enough to satisfy the hugepage allocation request. This is
clearly only a problem for large order allocations, of which hugepages
are the most obvious (to me).
Modify try_to_free_pages() to indicate how many pages were reclaimed.
Use that information in __alloc_pages() to eventually fail a large
__GFP_REPEAT allocation when we've reclaimed an order of pages equal to
or greater than the allocation's order. This relies on lumpy reclaim
functioning as advertised. Due to fragmentation, lumpy reclaim may not
be able to free up the order needed in one invocation, so multiple
iterations may be requred. In other words, the more fragmented memory
is, the more retry attempts __GFP_REPEAT will make (particularly for
higher order allocations).
This changes the semantics of __GFP_REPEAT subtly, but *only* for
allocations > PAGE_ALLOC_COSTLY_ORDER. With this patch, for those size
allocations, we will try up to some point (at least 1<<order reclaimed
pages), rather than forever (which is the case for allocations <=
PAGE_ALLOC_COSTLY_ORDER).
This change improves the /proc/sys/vm/nr_hugepages interface with a
follow-on patch that makes pool allocations use __GFP_REPEAT. Rather
than administrators repeatedly echo'ing a particular value into the
sysctl, and forcing reclaim into action manually, this change allows for
the sysctl to attempt a reasonable effort itself. Similarly, dynamic
pool growth should be more successful under load, as lumpy reclaim can
try to free up pages, rather than failing right away.
Choosing to reclaim only up to the order of the requested allocation
strikes a balance between not failing hugepage allocations and returning
to the caller when it's unlikely to every succeed. Because of lumpy
reclaim, if we have freed the order requested, hopefully it has been in
big chunks and those chunks will allow our allocation to succeed. If
that isn't the case after freeing up the current order, I don't think it
is likely to succeed in the future, although it is possible given a
particular fragmentation pattern.
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Tested-by: Mel Gorman <mel@csn.ul.ie>
Cc: Dave Hansen <haveblue@us.ibm.com>
Cc: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-29 07:58:25 +00:00
|
|
|
*
|
|
|
|
* returns: 0, if no pages reclaimed
|
|
|
|
* else, the number of pages reclaimed
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2008-04-28 09:12:12 +00:00
|
|
|
static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
|
2014-04-03 21:47:22 +00:00
|
|
|
struct scan_control *sc)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 23:26:06 +00:00
|
|
|
int initial_priority = sc->priority;
|
2017-05-03 21:55:03 +00:00
|
|
|
pg_data_t *last_pgdat;
|
|
|
|
struct zoneref *z;
|
|
|
|
struct zone *zone;
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 23:26:06 +00:00
|
|
|
retry:
|
2008-07-25 08:48:52 +00:00
|
|
|
delayacct_freepages_start();
|
|
|
|
|
2019-12-01 01:55:40 +00:00
|
|
|
if (!cgroup_reclaim(sc))
|
2016-07-28 22:46:59 +00:00
|
|
|
__count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2012-05-29 22:06:57 +00:00
|
|
|
do {
|
2022-07-14 06:49:18 +00:00
|
|
|
if (!sc->proactive)
|
|
|
|
vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
|
|
|
|
sc->priority);
|
2008-02-07 08:13:56 +00:00
|
|
|
sc->nr_scanned = 0;
|
2016-05-20 23:57:00 +00:00
|
|
|
shrink_zones(zonelist, sc);
|
2010-08-10 00:19:14 +00:00
|
|
|
|
2010-06-04 21:15:05 +00:00
|
|
|
if (sc->nr_reclaimed >= sc->nr_to_reclaim)
|
2014-08-06 23:06:12 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
if (sc->compaction_ready)
|
|
|
|
break;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2013-02-23 00:35:37 +00:00
|
|
|
/*
|
|
|
|
* If we're getting trouble reclaiming, start doing
|
|
|
|
* writepage even in laptop mode.
|
|
|
|
*/
|
|
|
|
if (sc->priority < DEF_PRIORITY - 2)
|
|
|
|
sc->may_writepage = 1;
|
2014-08-06 23:06:12 +00:00
|
|
|
} while (--sc->priority >= 0);
|
2010-06-04 21:15:05 +00:00
|
|
|
|
2017-05-03 21:55:03 +00:00
|
|
|
last_pgdat = NULL;
|
|
|
|
for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
|
|
|
|
sc->nodemask) {
|
|
|
|
if (zone->zone_pgdat == last_pgdat)
|
|
|
|
continue;
|
|
|
|
last_pgdat = zone->zone_pgdat;
|
2019-12-01 01:55:52 +00:00
|
|
|
|
2017-05-03 21:55:03 +00:00
|
|
|
snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
|
2019-12-01 01:55:52 +00:00
|
|
|
|
|
|
|
if (cgroup_reclaim(sc)) {
|
|
|
|
struct lruvec *lruvec;
|
|
|
|
|
|
|
|
lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
|
|
|
|
zone->zone_pgdat);
|
|
|
|
clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
|
|
|
|
}
|
2017-05-03 21:55:03 +00:00
|
|
|
}
|
|
|
|
|
2008-07-25 08:48:52 +00:00
|
|
|
delayacct_freepages_end();
|
|
|
|
|
2010-06-04 21:15:05 +00:00
|
|
|
if (sc->nr_reclaimed)
|
|
|
|
return sc->nr_reclaimed;
|
|
|
|
|
2012-01-13 01:19:49 +00:00
|
|
|
/* Aborted reclaim to try compaction? don't OOM, then */
|
2014-08-06 23:06:12 +00:00
|
|
|
if (sc->compaction_ready)
|
2012-01-13 01:19:33 +00:00
|
|
|
return 1;
|
|
|
|
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
/*
|
|
|
|
* We make inactive:active ratio decisions based on the node's
|
|
|
|
* composition of memory, but a restrictive reclaim_idx or a
|
|
|
|
* memory.low cgroup setting can exempt large amounts of
|
|
|
|
* memory from reclaim. Neither of which are very common, so
|
|
|
|
* instead of doing costly eligibility calculations of the
|
|
|
|
* entire cgroup subtree up front, we assume the estimates are
|
|
|
|
* good, and retry with forcible deactivation if that fails.
|
|
|
|
*/
|
|
|
|
if (sc->skipped_deactivate) {
|
|
|
|
sc->priority = initial_priority;
|
|
|
|
sc->force_deactivate = 1;
|
|
|
|
sc->skipped_deactivate = 0;
|
|
|
|
goto retry;
|
|
|
|
}
|
|
|
|
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 23:26:06 +00:00
|
|
|
/* Untapped cgroup reserves? Don't OOM, retry. */
|
mm/vmscan: more restrictive condition for retry in do_try_to_free_pages
By reviewing code, I find that when enter do_try_to_free_pages, the
may_thrash is always clear, and it will retry shrink zones to tap
cgroup's reserves memory by setting may_thrash when the former
shrink_zones reclaim nothing.
However, when memcg is disabled or on legacy hierarchy, or there do not
have any memcg protected by low limit, it should not do this useless
retry at all, for we do not have any cgroup's reserves memory to tap,
and we have already done hard work but made no progress, which as Michal
pointed out in former version, we are trying hard to control the retry
logical of page alloctor, and the current additional round of reclaim is
just lame.
Therefore, to avoid this unneeded retrying and make code more readable,
we remove the may_thrash field in scan_control, instead, introduce
memcg_low_reclaim and memcg_low_skipped, and only retry when
memcg_low_skipped, by setting memcg_low_reclaim.
[xieyisheng1@huawei.com: remove may_thrash field, introduce mem_cgroup_reclaim]
Link: http://lkml.kernel.org/r/1490191893-5923-1-git-send-email-ysxie@foxmail.com
Link: http://lkml.kernel.org/r/1490191893-5923-1-git-send-email-ysxie@foxmail.com
Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:57 +00:00
|
|
|
if (sc->memcg_low_skipped) {
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 23:26:06 +00:00
|
|
|
sc->priority = initial_priority;
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
sc->force_deactivate = 0;
|
mm/vmscan: more restrictive condition for retry in do_try_to_free_pages
By reviewing code, I find that when enter do_try_to_free_pages, the
may_thrash is always clear, and it will retry shrink zones to tap
cgroup's reserves memory by setting may_thrash when the former
shrink_zones reclaim nothing.
However, when memcg is disabled or on legacy hierarchy, or there do not
have any memcg protected by low limit, it should not do this useless
retry at all, for we do not have any cgroup's reserves memory to tap,
and we have already done hard work but made no progress, which as Michal
pointed out in former version, we are trying hard to control the retry
logical of page alloctor, and the current additional round of reclaim is
just lame.
Therefore, to avoid this unneeded retrying and make code more readable,
we remove the may_thrash field in scan_control, instead, introduce
memcg_low_reclaim and memcg_low_skipped, and only retry when
memcg_low_skipped, by setting memcg_low_reclaim.
[xieyisheng1@huawei.com: remove may_thrash field, introduce mem_cgroup_reclaim]
Link: http://lkml.kernel.org/r/1490191893-5923-1-git-send-email-ysxie@foxmail.com
Link: http://lkml.kernel.org/r/1490191893-5923-1-git-send-email-ysxie@foxmail.com
Signed-off-by: Yisheng Xie <xieyisheng1@huawei.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:57 +00:00
|
|
|
sc->memcg_low_reclaim = 1;
|
|
|
|
sc->memcg_low_skipped = 0;
|
mm: memcontrol: default hierarchy interface for memory
Introduce the basic control files to account, partition, and limit
memory using cgroups in default hierarchy mode.
This interface versioning allows us to address fundamental design
issues in the existing memory cgroup interface, further explained
below. The old interface will be maintained indefinitely, but a
clearer model and improved workload performance should encourage
existing users to switch over to the new one eventually.
The control files are thus:
- memory.current shows the current consumption of the cgroup and its
descendants, in bytes.
- memory.low configures the lower end of the cgroup's expected
memory consumption range. The kernel considers memory below that
boundary to be a reserve - the minimum that the workload needs in
order to make forward progress - and generally avoids reclaiming
it, unless there is an imminent risk of entering an OOM situation.
- memory.high configures the upper end of the cgroup's expected
memory consumption range. A cgroup whose consumption grows beyond
this threshold is forced into direct reclaim, to work off the
excess and to throttle new allocations heavily, but is generally
allowed to continue and the OOM killer is not invoked.
- memory.max configures the hard maximum amount of memory that the
cgroup is allowed to consume before the OOM killer is invoked.
- memory.events shows event counters that indicate how often the
cgroup was reclaimed while below memory.low, how often it was
forced to reclaim excess beyond memory.high, how often it hit
memory.max, and how often it entered OOM due to memory.max. This
allows users to identify configuration problems when observing a
degradation in workload performance. An overcommitted system will
have an increased rate of low boundary breaches, whereas increased
rates of high limit breaches, maximum hits, or even OOM situations
will indicate internally overcommitted cgroups.
For existing users of memory cgroups, the following deviations from
the current interface are worth pointing out and explaining:
- The original lower boundary, the soft limit, is defined as a limit
that is per default unset. As a result, the set of cgroups that
global reclaim prefers is opt-in, rather than opt-out. The costs
for optimizing these mostly negative lookups are so high that the
implementation, despite its enormous size, does not even provide
the basic desirable behavior. First off, the soft limit has no
hierarchical meaning. All configured groups are organized in a
global rbtree and treated like equal peers, regardless where they
are located in the hierarchy. This makes subtree delegation
impossible. Second, the soft limit reclaim pass is so aggressive
that it not just introduces high allocation latencies into the
system, but also impacts system performance due to overreclaim, to
the point where the feature becomes self-defeating.
The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries, which makes delegation
of subtrees possible. Secondly, new cgroups have no reserve per
default and in the common case most cgroups are eligible for the
preferred reclaim pass. This allows the new low boundary to be
efficiently implemented with just a minor addition to the generic
reclaim code, without the need for out-of-band data structures and
reclaim passes. Because the generic reclaim code considers all
cgroups except for the ones running low in the preferred first
reclaim pass, overreclaim of individual groups is eliminated as
well, resulting in much better overall workload performance.
- The original high boundary, the hard limit, is defined as a strict
limit that can not budge, even if the OOM killer has to be called.
But this generally goes against the goal of making the most out of
the available memory. The memory consumption of workloads varies
during runtime, and that requires users to overcommit. But doing
that with a strict upper limit requires either a fairly accurate
prediction of the working set size or adding slack to the limit.
Since working set size estimation is hard and error prone, and
getting it wrong results in OOM kills, most users tend to err on
the side of a looser limit and end up wasting precious resources.
The memory.high boundary on the other hand can be set much more
conservatively. When hit, it throttles allocations by forcing
them into direct reclaim to work off the excess, but it never
invokes the OOM killer. As a result, a high boundary that is
chosen too aggressively will not terminate the processes, but
instead it will lead to gradual performance degradation. The user
can monitor this and make corrections until the minimal memory
footprint that still gives acceptable performance is found.
In extreme cases, with many concurrent allocations and a complete
breakdown of reclaim progress within the group, the high boundary
can be exceeded. But even then it's mostly better to satisfy the
allocation from the slack available in other groups or the rest of
the system than killing the group. Otherwise, memory.max is there
to limit this type of spillover and ultimately contain buggy or
even malicious applications.
- The original control file names are unwieldy and inconsistent in
many different ways. For example, the upper boundary hit count is
exported in the memory.failcnt file, but an OOM event count has to
be manually counted by listening to memory.oom_control events, and
lower boundary / soft limit events have to be counted by first
setting a threshold for that value and then counting those events.
Also, usage and limit files encode their units in the filename.
That makes the filenames very long, even though this is not
information that a user needs to be reminded of every time they
type out those names.
To address these naming issues, as well as to signal clearly that
the new interface carries a new configuration model, the naming
conventions in it necessarily differ from the old interface.
- The original limit files indicate the state of an unset limit with
a very high number, and a configured limit can be unset by echoing
-1 into those files. But that very high number is implementation
and architecture dependent and not very descriptive. And while -1
can be understood as an underflow into the highest possible value,
-2 or -10M etc. do not work, so it's not inconsistent.
memory.low, memory.high, and memory.max will use the string
"infinity" to indicate and set the highest possible value.
[akpm@linux-foundation.org: use seq_puts() for basic strings]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 23:26:06 +00:00
|
|
|
goto retry;
|
|
|
|
}
|
|
|
|
|
2010-06-04 21:15:05 +00:00
|
|
|
return 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
static bool allow_direct_reclaim(pg_data_t *pgdat)
|
2012-07-31 23:44:35 +00:00
|
|
|
{
|
|
|
|
struct zone *zone;
|
|
|
|
unsigned long pfmemalloc_reserve = 0;
|
|
|
|
unsigned long free_pages = 0;
|
|
|
|
int i;
|
|
|
|
bool wmark_ok;
|
|
|
|
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
|
|
|
|
return true;
|
|
|
|
|
2012-07-31 23:44:35 +00:00
|
|
|
for (i = 0; i <= ZONE_NORMAL; i++) {
|
|
|
|
zone = &pgdat->node_zones[i];
|
2017-05-03 21:51:54 +00:00
|
|
|
if (!managed_zone(zone))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (!zone_reclaimable_pages(zone))
|
2014-06-04 23:07:35 +00:00
|
|
|
continue;
|
|
|
|
|
2012-07-31 23:44:35 +00:00
|
|
|
pfmemalloc_reserve += min_wmark_pages(zone);
|
|
|
|
free_pages += zone_page_state(zone, NR_FREE_PAGES);
|
|
|
|
}
|
|
|
|
|
2014-06-04 23:07:35 +00:00
|
|
|
/* If there are no reserves (unexpected config) then do not throttle */
|
|
|
|
if (!pfmemalloc_reserve)
|
|
|
|
return true;
|
|
|
|
|
2012-07-31 23:44:35 +00:00
|
|
|
wmark_ok = free_pages > pfmemalloc_reserve / 2;
|
|
|
|
|
|
|
|
/* kswapd must be awake if processes are being throttled */
|
|
|
|
if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
|
2020-06-03 22:59:01 +00:00
|
|
|
if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
|
|
|
|
WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
|
2020-04-02 04:10:12 +00:00
|
|
|
|
2012-07-31 23:44:35 +00:00
|
|
|
wake_up_interruptible(&pgdat->kswapd_wait);
|
|
|
|
}
|
|
|
|
|
|
|
|
return wmark_ok;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Throttle direct reclaimers if backing storage is backed by the network
|
|
|
|
* and the PFMEMALLOC reserve for the preferred node is getting dangerously
|
|
|
|
* depleted. kswapd will continue to make progress and wake the processes
|
2012-11-27 00:29:48 +00:00
|
|
|
* when the low watermark is reached.
|
|
|
|
*
|
|
|
|
* Returns true if a fatal signal was delivered during throttling. If this
|
|
|
|
* happens, the page allocator should not consider triggering the OOM killer.
|
2012-07-31 23:44:35 +00:00
|
|
|
*/
|
2012-11-27 00:29:48 +00:00
|
|
|
static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
|
2012-07-31 23:44:35 +00:00
|
|
|
nodemask_t *nodemask)
|
|
|
|
{
|
2014-06-04 23:07:35 +00:00
|
|
|
struct zoneref *z;
|
2012-07-31 23:44:35 +00:00
|
|
|
struct zone *zone;
|
2014-06-04 23:07:35 +00:00
|
|
|
pg_data_t *pgdat = NULL;
|
2012-07-31 23:44:35 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Kernel threads should not be throttled as they may be indirectly
|
|
|
|
* responsible for cleaning pages necessary for reclaim to make forward
|
|
|
|
* progress. kjournald for example may enter direct reclaim while
|
|
|
|
* committing a transaction where throttling it could forcing other
|
|
|
|
* processes to block on log_wait_commit().
|
|
|
|
*/
|
|
|
|
if (current->flags & PF_KTHREAD)
|
2012-11-27 00:29:48 +00:00
|
|
|
goto out;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If a fatal signal is pending, this process should not throttle.
|
|
|
|
* It should return quickly so it can exit and free its memory
|
|
|
|
*/
|
|
|
|
if (fatal_signal_pending(current))
|
|
|
|
goto out;
|
2012-07-31 23:44:35 +00:00
|
|
|
|
2014-06-04 23:07:35 +00:00
|
|
|
/*
|
|
|
|
* Check if the pfmemalloc reserves are ok by finding the first node
|
|
|
|
* with a usable ZONE_NORMAL or lower zone. The expectation is that
|
|
|
|
* GFP_KERNEL will be required for allocating network buffers when
|
|
|
|
* swapping over the network so ZONE_HIGHMEM is unusable.
|
|
|
|
*
|
|
|
|
* Throttling is based on the first usable node and throttled processes
|
|
|
|
* wait on a queue until kswapd makes progress and wakes them. There
|
|
|
|
* is an affinity then between processes waking up and where reclaim
|
|
|
|
* progress has been made assuming the process wakes on the same node.
|
|
|
|
* More importantly, processes running on remote nodes will not compete
|
|
|
|
* for remote pfmemalloc reserves and processes on different nodes
|
|
|
|
* should make reasonable progress.
|
|
|
|
*/
|
|
|
|
for_each_zone_zonelist_nodemask(zone, z, zonelist,
|
2015-01-26 20:58:41 +00:00
|
|
|
gfp_zone(gfp_mask), nodemask) {
|
2014-06-04 23:07:35 +00:00
|
|
|
if (zone_idx(zone) > ZONE_NORMAL)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/* Throttle based on the first usable node */
|
|
|
|
pgdat = zone->zone_pgdat;
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
if (allow_direct_reclaim(pgdat))
|
2014-06-04 23:07:35 +00:00
|
|
|
goto out;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* If no zone was usable by the allocation flags then do not throttle */
|
|
|
|
if (!pgdat)
|
2012-11-27 00:29:48 +00:00
|
|
|
goto out;
|
2012-07-31 23:44:35 +00:00
|
|
|
|
2012-07-31 23:44:39 +00:00
|
|
|
/* Account for the throttling */
|
|
|
|
count_vm_event(PGSCAN_DIRECT_THROTTLE);
|
|
|
|
|
2012-07-31 23:44:35 +00:00
|
|
|
/*
|
|
|
|
* If the caller cannot enter the filesystem, it's possible that it
|
|
|
|
* is due to the caller holding an FS lock or performing a journal
|
|
|
|
* transaction in the case of a filesystem like ext[3|4]. In this case,
|
|
|
|
* it is not safe to block on pfmemalloc_wait as kswapd could be
|
|
|
|
* blocked waiting on the same lock. Instead, throttle for up to a
|
|
|
|
* second before continuing.
|
|
|
|
*/
|
2021-09-02 21:59:50 +00:00
|
|
|
if (!(gfp_mask & __GFP_FS))
|
2012-07-31 23:44:35 +00:00
|
|
|
wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
allow_direct_reclaim(pgdat), HZ);
|
2021-09-02 21:59:50 +00:00
|
|
|
else
|
|
|
|
/* Throttle until kswapd wakes the process */
|
|
|
|
wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
|
|
|
|
allow_direct_reclaim(pgdat));
|
2012-11-27 00:29:48 +00:00
|
|
|
|
|
|
|
if (fatal_signal_pending(current))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
out:
|
|
|
|
return false;
|
2012-07-31 23:44:35 +00:00
|
|
|
}
|
|
|
|
|
2008-04-28 09:12:12 +00:00
|
|
|
unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
|
2009-03-31 22:23:31 +00:00
|
|
|
gfp_t gfp_mask, nodemask_t *nodemask)
|
2008-02-07 08:13:56 +00:00
|
|
|
{
|
2010-08-10 00:19:16 +00:00
|
|
|
unsigned long nr_reclaimed;
|
2008-02-07 08:13:56 +00:00
|
|
|
struct scan_control sc = {
|
2014-08-06 23:06:19 +00:00
|
|
|
.nr_to_reclaim = SWAP_CLUSTER_MAX,
|
2017-07-06 22:36:50 +00:00
|
|
|
.gfp_mask = current_gfp_context(gfp_mask),
|
2016-07-28 22:45:37 +00:00
|
|
|
.reclaim_idx = gfp_zone(gfp_mask),
|
2014-08-06 23:06:19 +00:00
|
|
|
.order = order,
|
|
|
|
.nodemask = nodemask,
|
|
|
|
.priority = DEF_PRIORITY,
|
2008-02-07 08:13:56 +00:00
|
|
|
.may_writepage = !laptop_mode,
|
2009-03-31 22:19:30 +00:00
|
|
|
.may_unmap = 1,
|
2009-04-21 19:24:57 +00:00
|
|
|
.may_swap = 1,
|
2008-02-07 08:13:56 +00:00
|
|
|
};
|
|
|
|
|
2018-08-17 22:45:19 +00:00
|
|
|
/*
|
|
|
|
* scan_control uses s8 fields for order, priority, and reclaim_idx.
|
|
|
|
* Confirm they are large enough for max values.
|
|
|
|
*/
|
|
|
|
BUILD_BUG_ON(MAX_ORDER > S8_MAX);
|
|
|
|
BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
|
|
|
|
BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
|
|
|
|
|
2012-07-31 23:44:35 +00:00
|
|
|
/*
|
2012-11-27 00:29:48 +00:00
|
|
|
* Do not enter reclaim if fatal signal was delivered while throttled.
|
|
|
|
* 1 is returned so that the page allocator does not OOM kill at this
|
|
|
|
* point.
|
2012-07-31 23:44:35 +00:00
|
|
|
*/
|
2017-07-06 22:36:50 +00:00
|
|
|
if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
|
2012-07-31 23:44:35 +00:00
|
|
|
return 1;
|
|
|
|
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(current, &sc.reclaim_state);
|
2019-05-14 00:19:14 +00:00
|
|
|
trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
|
2010-08-10 00:19:16 +00:00
|
|
|
|
2014-04-03 21:47:22 +00:00
|
|
|
nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
|
2010-08-10 00:19:16 +00:00
|
|
|
|
|
|
|
trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(current, NULL);
|
2010-08-10 00:19:16 +00:00
|
|
|
|
|
|
|
return nr_reclaimed;
|
2008-02-07 08:13:56 +00:00
|
|
|
}
|
|
|
|
|
2012-07-31 23:43:02 +00:00
|
|
|
#ifdef CONFIG_MEMCG
|
2008-02-07 08:13:56 +00:00
|
|
|
|
2019-08-30 23:04:50 +00:00
|
|
|
/* Only used by soft limit reclaim. Do not reuse for anything else. */
|
2016-07-28 22:46:02 +00:00
|
|
|
unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
|
2009-09-23 22:56:39 +00:00
|
|
|
gfp_t gfp_mask, bool noswap,
|
2016-07-28 22:46:05 +00:00
|
|
|
pg_data_t *pgdat,
|
2011-05-26 23:25:25 +00:00
|
|
|
unsigned long *nr_scanned)
|
2009-09-23 22:56:39 +00:00
|
|
|
{
|
2019-12-01 01:55:46 +00:00
|
|
|
struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
|
2009-09-23 22:56:39 +00:00
|
|
|
struct scan_control sc = {
|
2010-08-11 01:03:02 +00:00
|
|
|
.nr_to_reclaim = SWAP_CLUSTER_MAX,
|
2014-08-06 23:06:19 +00:00
|
|
|
.target_mem_cgroup = memcg,
|
2009-09-23 22:56:39 +00:00
|
|
|
.may_writepage = !laptop_mode,
|
|
|
|
.may_unmap = 1,
|
2016-07-28 22:45:37 +00:00
|
|
|
.reclaim_idx = MAX_NR_ZONES - 1,
|
2009-09-23 22:56:39 +00:00
|
|
|
.may_swap = !noswap,
|
|
|
|
};
|
2011-05-26 23:25:25 +00:00
|
|
|
|
2019-08-30 23:04:50 +00:00
|
|
|
WARN_ON_ONCE(!current->reclaim_state);
|
|
|
|
|
2009-09-23 22:56:39 +00:00
|
|
|
sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
|
|
|
|
(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
|
2010-08-10 00:19:56 +00:00
|
|
|
|
2012-05-29 22:06:57 +00:00
|
|
|
trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
|
2019-05-14 00:19:14 +00:00
|
|
|
sc.gfp_mask);
|
2010-08-10 00:19:56 +00:00
|
|
|
|
2009-09-23 22:56:39 +00:00
|
|
|
/*
|
|
|
|
* NOTE: Although we can get the priority field, using it
|
|
|
|
* here is not a good idea, since it limits the pages we can scan.
|
2016-07-28 22:46:02 +00:00
|
|
|
* if we don't reclaim here, the shrink_node from balance_pgdat
|
2009-09-23 22:56:39 +00:00
|
|
|
* will pick up pages from other mem cgroup's as well. We hack
|
|
|
|
* the priority and make it zero.
|
|
|
|
*/
|
2019-12-01 01:55:46 +00:00
|
|
|
shrink_lruvec(lruvec, &sc);
|
2010-08-10 00:19:56 +00:00
|
|
|
|
|
|
|
trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
|
|
|
|
|
2011-05-26 23:25:25 +00:00
|
|
|
*nr_scanned = sc.nr_scanned;
|
2019-07-16 23:26:12 +00:00
|
|
|
|
2009-09-23 22:56:39 +00:00
|
|
|
return sc.nr_reclaimed;
|
|
|
|
}
|
|
|
|
|
2012-01-13 01:18:32 +00:00
|
|
|
unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
|
2014-10-09 22:28:56 +00:00
|
|
|
unsigned long nr_pages,
|
2009-01-08 02:08:24 +00:00
|
|
|
gfp_t gfp_mask,
|
2022-12-16 09:46:33 +00:00
|
|
|
unsigned int reclaim_options)
|
2008-02-07 08:13:56 +00:00
|
|
|
{
|
2010-08-10 00:19:56 +00:00
|
|
|
unsigned long nr_reclaimed;
|
2017-05-08 22:59:50 +00:00
|
|
|
unsigned int noreclaim_flag;
|
2008-02-07 08:13:56 +00:00
|
|
|
struct scan_control sc = {
|
2014-10-09 22:28:56 +00:00
|
|
|
.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
|
mm: introduce memalloc_nofs_{save,restore} API
GFP_NOFS context is used for the following 5 reasons currently:
- to prevent from deadlocks when the lock held by the allocation
context would be needed during the memory reclaim
- to prevent from stack overflows during the reclaim because the
allocation is performed from a deep context already
- to prevent lockups when the allocation context depends on other
reclaimers to make a forward progress indirectly
- just in case because this would be safe from the fs POV
- silence lockdep false positives
Unfortunately overuse of this allocation context brings some problems to
the MM. Memory reclaim is much weaker (especially during heavy FS
metadata workloads), OOM killer cannot be invoked because the MM layer
doesn't have enough information about how much memory is freeable by the
FS layer.
In many cases it is far from clear why the weaker context is even used
and so it might be used unnecessarily. We would like to get rid of
those as much as possible. One way to do that is to use the flag in
scopes rather than isolated cases. Such a scope is declared when really
necessary, tracked per task and all the allocation requests from within
the context will simply inherit the GFP_NOFS semantic.
Not only this is easier to understand and maintain because there are
much less problematic contexts than specific allocation requests, this
also helps code paths where FS layer interacts with other layers (e.g.
crypto, security modules, MM etc...) and there is no easy way to convey
the allocation context between the layers.
Introduce memalloc_nofs_{save,restore} API to control the scope of
GFP_NOFS allocation context. This is basically copying
memalloc_noio_{save,restore} API we have for other restricted allocation
context GFP_NOIO. The PF_MEMALLOC_NOFS flag already exists and it is
just an alias for PF_FSTRANS which has been xfs specific until recently.
There are no more PF_FSTRANS users anymore so let's just drop it.
PF_MEMALLOC_NOFS is now checked in the MM layer and drops __GFP_FS
implicitly same as PF_MEMALLOC_NOIO drops __GFP_IO. memalloc_noio_flags
is renamed to current_gfp_context because it now cares about both
PF_MEMALLOC_NOFS and PF_MEMALLOC_NOIO contexts. Xfs code paths preserve
their semantic. kmem_flags_convert() doesn't need to evaluate the flag
anymore.
This patch shouldn't introduce any functional changes.
Let's hope that filesystems will drop direct GFP_NOFS (resp. ~__GFP_FS)
usage as much as possible and only use a properly documented
memalloc_nofs_{save,restore} checkpoints where they are appropriate.
[akpm@linux-foundation.org: fix comment typo, reflow comment]
Link: http://lkml.kernel.org/r/20170306131408.9828-5-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Chris Mason <clm@fb.com>
Cc: David Sterba <dsterba@suse.cz>
Cc: Jan Kara <jack@suse.cz>
Cc: Brian Foster <bfoster@redhat.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: Nikolay Borisov <nborisov@suse.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:15 +00:00
|
|
|
.gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
|
2011-05-25 00:12:26 +00:00
|
|
|
(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
|
2016-07-28 22:45:37 +00:00
|
|
|
.reclaim_idx = MAX_NR_ZONES - 1,
|
2014-08-06 23:06:19 +00:00
|
|
|
.target_mem_cgroup = memcg,
|
|
|
|
.priority = DEF_PRIORITY,
|
|
|
|
.may_writepage = !laptop_mode,
|
|
|
|
.may_unmap = 1,
|
2022-07-14 06:49:18 +00:00
|
|
|
.may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
|
|
|
|
.proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
|
2011-05-25 00:12:26 +00:00
|
|
|
};
|
2011-05-26 23:25:33 +00:00
|
|
|
/*
|
2019-12-01 01:50:16 +00:00
|
|
|
* Traverse the ZONELIST_FALLBACK zonelist of the current node to put
|
|
|
|
* equal pressure on all the nodes. This is based on the assumption that
|
|
|
|
* the reclaim does not bail out early.
|
2011-05-26 23:25:33 +00:00
|
|
|
*/
|
2019-12-01 01:50:16 +00:00
|
|
|
struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
|
2011-05-26 23:25:33 +00:00
|
|
|
|
2019-12-01 01:50:16 +00:00
|
|
|
set_task_reclaim_state(current, &sc.reclaim_state);
|
2019-05-14 00:19:14 +00:00
|
|
|
trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
|
2017-05-08 22:59:50 +00:00
|
|
|
noreclaim_flag = memalloc_noreclaim_save();
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:06:27 +00:00
|
|
|
|
2014-04-03 21:47:22 +00:00
|
|
|
nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:06:27 +00:00
|
|
|
|
2017-05-08 22:59:50 +00:00
|
|
|
memalloc_noreclaim_restore(noreclaim_flag);
|
2010-08-10 00:19:56 +00:00
|
|
|
trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(current, NULL);
|
2010-08-10 00:19:56 +00:00
|
|
|
|
|
|
|
return nr_reclaimed;
|
2008-02-07 08:13:56 +00:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
|
2012-01-13 01:17:52 +00:00
|
|
|
{
|
2012-01-13 01:18:06 +00:00
|
|
|
struct mem_cgroup *memcg;
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
struct lruvec *lruvec;
|
2012-01-13 01:17:52 +00:00
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
if (lru_gen_enabled()) {
|
|
|
|
lru_gen_age_node(pgdat, sc);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2021-09-02 21:59:23 +00:00
|
|
|
if (!can_age_anon_pages(pgdat, sc))
|
2012-01-13 01:18:06 +00:00
|
|
|
return;
|
|
|
|
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
lruvec = mem_cgroup_lruvec(NULL, pgdat);
|
|
|
|
if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
|
|
|
|
return;
|
|
|
|
|
2012-01-13 01:18:06 +00:00
|
|
|
memcg = mem_cgroup_iter(NULL, NULL, NULL);
|
|
|
|
do {
|
mm: vmscan: enforce inactive:active ratio at the reclaim root
We split the LRU lists into inactive and an active parts to maximize
workingset protection while allowing just enough inactive cache space to
faciltate readahead and writeback for one-off file accesses (e.g. a
linear scan through a file, or logging); or just enough inactive anon to
maintain recent reference information when reclaim needs to swap.
With cgroups and their nested LRU lists, we currently don't do this
correctly. While recursive cgroup reclaim establishes a relative LRU
order among the pages of all involved cgroups, inactive:active size
decisions are done on a per-cgroup level. As a result, we'll reclaim a
cgroup's workingset when it doesn't have cold pages, even when one of its
siblings has plenty of it that should be reclaimed first.
For example: workload A has 50M worth of hot cache but doesn't do any
one-off file accesses; meanwhile, parallel workload B scans files and
rarely accesses the same page twice.
If these workloads were to run in an uncgrouped system, A would be
protected from the high rate of cache faults from B. But if they were put
in parallel cgroups for memory accounting purposes, B's fast cache fault
rate would push out the hot cache pages of A. This is unexpected and
undesirable - the "scan resistance" of the page cache is broken.
This patch moves inactive:active size balancing decisions to the root of
reclaim - the same level where the LRU order is established.
It does this by looking at the recursive size of the inactive and the
active file sets of the cgroup subtree at the beginning of the reclaim
cycle, and then making a decision - scan or skip active pages - that
applies throughout the entire run and to every cgroup involved.
With that in place, in the test above, the VM will recognize that there
are plenty of inactive pages in the combined cache set of workloads A and
B and prefer the one-off cache in B over the hot pages in A. The scan
resistance of the cache is restored.
[cai@lca.pw: fix some -Wenum-conversion warnings]
Link: http://lkml.kernel.org/r/1573848697-29262-1-git-send-email-cai@lca.pw
Link: http://lkml.kernel.org/r/20191107205334.158354-4-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 01:56:02 +00:00
|
|
|
lruvec = mem_cgroup_lruvec(memcg, pgdat);
|
|
|
|
shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
|
|
|
|
sc, LRU_ACTIVE_ANON);
|
2012-01-13 01:18:06 +00:00
|
|
|
memcg = mem_cgroup_iter(NULL, memcg, NULL);
|
|
|
|
} while (memcg);
|
2012-01-13 01:17:52 +00:00
|
|
|
}
|
|
|
|
|
2020-06-03 22:59:01 +00:00
|
|
|
static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
{
|
|
|
|
int i;
|
|
|
|
struct zone *zone;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Check for watermark boosts top-down as the higher zones
|
|
|
|
* are more likely to be boosted. Both watermarks and boosts
|
2020-08-12 01:33:26 +00:00
|
|
|
* should not be checked at the same time as reclaim would
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
* start prematurely when there is no boosting and a lower
|
|
|
|
* zone is balanced.
|
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
for (i = highest_zoneidx; i >= 0; i--) {
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
zone = pgdat->node_zones + i;
|
|
|
|
if (!managed_zone(zone))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (zone->watermark_boost)
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
/*
|
|
|
|
* Returns true if there is an eligible zone balanced for the request order
|
2020-06-03 22:59:01 +00:00
|
|
|
* and highest_zoneidx
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
|
2012-11-29 21:54:23 +00:00
|
|
|
{
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
int i;
|
|
|
|
unsigned long mark = -1;
|
|
|
|
struct zone *zone;
|
2012-11-29 21:54:23 +00:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
/*
|
|
|
|
* Check watermarks bottom-up as lower zones are more likely to
|
|
|
|
* meet watermarks.
|
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
for (i = 0; i <= highest_zoneidx; i++) {
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
zone = pgdat->node_zones + i;
|
2016-07-28 22:45:56 +00:00
|
|
|
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
if (!managed_zone(zone))
|
|
|
|
continue;
|
|
|
|
|
NUMA balancing: optimize page placement for memory tiering system
With the advent of various new memory types, some machines will have
multiple types of memory, e.g. DRAM and PMEM (persistent memory). The
memory subsystem of these machines can be called memory tiering system,
because the performance of the different types of memory are usually
different.
In such system, because of the memory accessing pattern changing etc,
some pages in the slow memory may become hot globally. So in this
patch, the NUMA balancing mechanism is enhanced to optimize the page
placement among the different memory types according to hot/cold
dynamically.
In a typical memory tiering system, there are CPUs, fast memory and slow
memory in each physical NUMA node. The CPUs and the fast memory will be
put in one logical node (called fast memory node), while the slow memory
will be put in another (faked) logical node (called slow memory node).
That is, the fast memory is regarded as local while the slow memory is
regarded as remote. So it's possible for the recently accessed pages in
the slow memory node to be promoted to the fast memory node via the
existing NUMA balancing mechanism.
The original NUMA balancing mechanism will stop to migrate pages if the
free memory of the target node becomes below the high watermark. This
is a reasonable policy if there's only one memory type. But this makes
the original NUMA balancing mechanism almost do not work to optimize
page placement among different memory types. Details are as follows.
It's the common cases that the working-set size of the workload is
larger than the size of the fast memory nodes. Otherwise, it's
unnecessary to use the slow memory at all. So, there are almost always
no enough free pages in the fast memory nodes, so that the globally hot
pages in the slow memory node cannot be promoted to the fast memory
node. To solve the issue, we have 2 choices as follows,
a. Ignore the free pages watermark checking when promoting hot pages
from the slow memory node to the fast memory node. This will
create some memory pressure in the fast memory node, thus trigger
the memory reclaiming. So that, the cold pages in the fast memory
node will be demoted to the slow memory node.
b. Define a new watermark called wmark_promo which is higher than
wmark_high, and have kswapd reclaiming pages until free pages reach
such watermark. The scenario is as follows: when we want to promote
hot-pages from a slow memory to a fast memory, but fast memory's free
pages would go lower than high watermark with such promotion, we wake
up kswapd with wmark_promo watermark in order to demote cold pages and
free us up some space. So, next time we want to promote hot-pages we
might have a chance of doing so.
The choice "a" may create high memory pressure in the fast memory node.
If the memory pressure of the workload is high, the memory pressure
may become so high that the memory allocation latency of the workload
is influenced, e.g. the direct reclaiming may be triggered.
The choice "b" works much better at this aspect. If the memory
pressure of the workload is high, the hot pages promotion will stop
earlier because its allocation watermark is higher than that of the
normal memory allocation. So in this patch, choice "b" is implemented.
A new zone watermark (WMARK_PROMO) is added. Which is larger than the
high watermark and can be controlled via watermark_scale_factor.
In addition to the original page placement optimization among sockets,
the NUMA balancing mechanism is extended to be used to optimize page
placement according to hot/cold among different memory types. So the
sysctl user space interface (numa_balancing) is extended in a backward
compatible way as follow, so that the users can enable/disable these
functionality individually.
The sysctl is converted from a Boolean value to a bits field. The
definition of the flags is,
- 0: NUMA_BALANCING_DISABLED
- 1: NUMA_BALANCING_NORMAL
- 2: NUMA_BALANCING_MEMORY_TIERING
We have tested the patch with the pmbench memory accessing benchmark
with the 80:20 read/write ratio and the Gauss access address
distribution on a 2 socket Intel server with Optane DC Persistent
Memory Model. The test results shows that the pmbench score can
improve up to 95.9%.
Thanks Andrew Morton to help fix the document format error.
Link: https://lkml.kernel.org/r/20220221084529.1052339-3-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: zhongjiang-ali <zhongjiang-ali@linux.alibaba.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Feng Tang <feng.tang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-22 21:46:23 +00:00
|
|
|
if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
|
|
|
|
mark = wmark_pages(zone, WMARK_PROMO);
|
|
|
|
else
|
|
|
|
mark = high_wmark_pages(zone);
|
2020-06-03 22:59:01 +00:00
|
|
|
if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2022-04-29 06:16:03 +00:00
|
|
|
* If a node has no managed zone within highest_zoneidx, it does not
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
* need balancing by definition. This can happen if a zone-restricted
|
|
|
|
* allocation tries to wake a remote kswapd.
|
|
|
|
*/
|
|
|
|
if (mark == -1)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
return false;
|
2012-11-29 21:54:23 +00:00
|
|
|
}
|
|
|
|
|
mm, vmscan: only clear pgdat congested/dirty/writeback state when balanced
A pgdat tracks if recent reclaim encountered too many dirty, writeback
or congested pages. The flags control whether kswapd writes pages back
from reclaim context, tags pages for immediate reclaim when IO
completes, whether processes block on wait_iff_congested and whether
kswapd blocks when too many pages marked for immediate reclaim are
encountered.
The state is cleared in a check function with side-effects. With the
patch "mm, vmscan: fix zone balance check in prepare_kswapd_sleep", the
timing of when the bits get cleared changed. Due to the way the check
works, it'll clear the bits if ZONE_DMA is balanced for a GFP_DMA
allocation because it does not account for lowmem reserves properly.
For the simoop workload, kswapd is not stalling when it should due to
the premature clearing, writing pages from reclaim context like crazy
and generally being unhelpful.
This patch resets the pgdat bits related to page reclaim only when
kswapd is going to sleep. The comparison with simoop is then
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla fixcheck-v2 clear-v2
Amean p50-Read 21670074.18 ( 0.00%) 20464344.18 ( 5.56%) 19786774.76 ( 8.69%)
Amean p95-Read 25456267.64 ( 0.00%) 25721423.64 ( -1.04%) 24101956.27 ( 5.32%)
Amean p99-Read 29369064.73 ( 0.00%) 30174230.76 ( -2.74%) 27691872.71 ( 5.71%)
Amean p50-Write 1390.30 ( 0.00%) 1395.28 ( -0.36%) 1011.91 ( 27.22%)
Amean p95-Write 412901.57 ( 0.00%) 37737.74 ( 90.86%) 34874.98 ( 91.55%)
Amean p99-Write 6668722.09 ( 0.00%) 666489.04 ( 90.01%) 575449.60 ( 91.37%)
Amean p50-Allocation 78714.31 ( 0.00%) 86286.22 ( -9.62%) 84246.26 ( -7.03%)
Amean p95-Allocation 175533.51 ( 0.00%) 351812.27 (-100.42%) 400058.43 (-127.91%)
Amean p99-Allocation 247003.02 ( 0.00%) 6291171.56 (-2447.00%) 10905600.00 (-4315.17%)
Read latency is improved, write latency is mostly improved but
allocation latency is regressed. kswapd is still reclaiming
inefficiently, pages are being written back from writeback context and a
host of other issues. However, given the change, it needed to be
spelled out why the side-effect was moved.
Link: http://lkml.kernel.org/r/20170309075657.25121-3-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:41 +00:00
|
|
|
/* Clear pgdat state for congested, dirty or under writeback. */
|
|
|
|
static void clear_pgdat_congested(pg_data_t *pgdat)
|
|
|
|
{
|
2019-12-01 01:55:52 +00:00
|
|
|
struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
|
|
|
|
|
|
|
|
clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
|
mm, vmscan: only clear pgdat congested/dirty/writeback state when balanced
A pgdat tracks if recent reclaim encountered too many dirty, writeback
or congested pages. The flags control whether kswapd writes pages back
from reclaim context, tags pages for immediate reclaim when IO
completes, whether processes block on wait_iff_congested and whether
kswapd blocks when too many pages marked for immediate reclaim are
encountered.
The state is cleared in a check function with side-effects. With the
patch "mm, vmscan: fix zone balance check in prepare_kswapd_sleep", the
timing of when the bits get cleared changed. Due to the way the check
works, it'll clear the bits if ZONE_DMA is balanced for a GFP_DMA
allocation because it does not account for lowmem reserves properly.
For the simoop workload, kswapd is not stalling when it should due to
the premature clearing, writing pages from reclaim context like crazy
and generally being unhelpful.
This patch resets the pgdat bits related to page reclaim only when
kswapd is going to sleep. The comparison with simoop is then
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla fixcheck-v2 clear-v2
Amean p50-Read 21670074.18 ( 0.00%) 20464344.18 ( 5.56%) 19786774.76 ( 8.69%)
Amean p95-Read 25456267.64 ( 0.00%) 25721423.64 ( -1.04%) 24101956.27 ( 5.32%)
Amean p99-Read 29369064.73 ( 0.00%) 30174230.76 ( -2.74%) 27691872.71 ( 5.71%)
Amean p50-Write 1390.30 ( 0.00%) 1395.28 ( -0.36%) 1011.91 ( 27.22%)
Amean p95-Write 412901.57 ( 0.00%) 37737.74 ( 90.86%) 34874.98 ( 91.55%)
Amean p99-Write 6668722.09 ( 0.00%) 666489.04 ( 90.01%) 575449.60 ( 91.37%)
Amean p50-Allocation 78714.31 ( 0.00%) 86286.22 ( -9.62%) 84246.26 ( -7.03%)
Amean p95-Allocation 175533.51 ( 0.00%) 351812.27 (-100.42%) 400058.43 (-127.91%)
Amean p99-Allocation 247003.02 ( 0.00%) 6291171.56 (-2447.00%) 10905600.00 (-4315.17%)
Read latency is improved, write latency is mostly improved but
allocation latency is regressed. kswapd is still reclaiming
inefficiently, pages are being written back from writeback context and a
host of other issues. However, given the change, it needed to be
spelled out why the side-effect was moved.
Link: http://lkml.kernel.org/r/20170309075657.25121-3-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:41 +00:00
|
|
|
clear_bit(PGDAT_DIRTY, &pgdat->flags);
|
|
|
|
clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
|
|
|
|
}
|
|
|
|
|
2012-07-31 23:44:35 +00:00
|
|
|
/*
|
|
|
|
* Prepare kswapd for sleeping. This verifies that there are no processes
|
|
|
|
* waiting in throttle_direct_reclaim() and that watermarks have been met.
|
|
|
|
*
|
|
|
|
* Returns true if kswapd is ready to sleep
|
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
|
|
|
|
int highest_zoneidx)
|
2009-12-15 01:58:53 +00:00
|
|
|
{
|
2012-07-31 23:44:35 +00:00
|
|
|
/*
|
mm, vmscan: prevent kswapd livelock due to pfmemalloc-throttled process being killed
Charles Shirron and Paul Cassella from Cray Inc have reported kswapd
stuck in a busy loop with nothing left to balance, but
kswapd_try_to_sleep() failing to sleep. Their analysis found the cause
to be a combination of several factors:
1. A process is waiting in throttle_direct_reclaim() on pgdat->pfmemalloc_wait
2. The process has been killed (by OOM in this case), but has not yet been
scheduled to remove itself from the waitqueue and die.
3. kswapd checks for throttled processes in prepare_kswapd_sleep():
if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
wake_up(&pgdat->pfmemalloc_wait);
return false; // kswapd will not go to sleep
}
However, for a process that was already killed, wake_up() does not remove
the process from the waitqueue, since try_to_wake_up() checks its state
first and returns false when the process is no longer waiting.
4. kswapd is running on the same CPU as the only CPU that the process is
allowed to run on (through cpus_allowed, or possibly single-cpu system).
5. CONFIG_PREEMPT_NONE=y kernel is used. If there's nothing to balance, kswapd
encounters no voluntary preemption points and repeatedly fails
prepare_kswapd_sleep(), blocking the process from running and removing
itself from the waitqueue, which would let kswapd sleep.
So, the source of the problem is that we prevent kswapd from going to
sleep until there are processes waiting on the pfmemalloc_wait queue,
and a process waiting on a queue is guaranteed to be removed from the
queue only when it gets scheduled. This was done to make sure that no
process is left sleeping on pfmemalloc_wait when kswapd itself goes to
sleep.
However, it isn't necessary to postpone kswapd sleep until the
pfmemalloc_wait queue actually empties. To prevent processes from being
left sleeping, it's actually enough to guarantee that all processes
waiting on pfmemalloc_wait queue have been woken up by the time we put
kswapd to sleep.
This patch therefore fixes this issue by substituting 'wake_up' with
'wake_up_all' and removing 'return false' in the code snippet from
prepare_kswapd_sleep() above. Note that if any process puts itself in
the queue after this waitqueue_active() check, or after the wake up
itself, it means that the process will also wake up kswapd - and since
we are under prepare_to_wait(), the wake up won't be missed. Also we
update the comment prepare_kswapd_sleep() to hopefully more clearly
describe the races it is preventing.
Fixes: 5515061d22f0 ("mm: throttle direct reclaimers if PF_MEMALLOC reserves are low and swap is backed by network storage")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: <stable@vger.kernel.org> [3.6+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-01-08 22:32:40 +00:00
|
|
|
* The throttled processes are normally woken up in balance_pgdat() as
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
* soon as allow_direct_reclaim() is true. But there is a potential
|
mm, vmscan: prevent kswapd livelock due to pfmemalloc-throttled process being killed
Charles Shirron and Paul Cassella from Cray Inc have reported kswapd
stuck in a busy loop with nothing left to balance, but
kswapd_try_to_sleep() failing to sleep. Their analysis found the cause
to be a combination of several factors:
1. A process is waiting in throttle_direct_reclaim() on pgdat->pfmemalloc_wait
2. The process has been killed (by OOM in this case), but has not yet been
scheduled to remove itself from the waitqueue and die.
3. kswapd checks for throttled processes in prepare_kswapd_sleep():
if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
wake_up(&pgdat->pfmemalloc_wait);
return false; // kswapd will not go to sleep
}
However, for a process that was already killed, wake_up() does not remove
the process from the waitqueue, since try_to_wake_up() checks its state
first and returns false when the process is no longer waiting.
4. kswapd is running on the same CPU as the only CPU that the process is
allowed to run on (through cpus_allowed, or possibly single-cpu system).
5. CONFIG_PREEMPT_NONE=y kernel is used. If there's nothing to balance, kswapd
encounters no voluntary preemption points and repeatedly fails
prepare_kswapd_sleep(), blocking the process from running and removing
itself from the waitqueue, which would let kswapd sleep.
So, the source of the problem is that we prevent kswapd from going to
sleep until there are processes waiting on the pfmemalloc_wait queue,
and a process waiting on a queue is guaranteed to be removed from the
queue only when it gets scheduled. This was done to make sure that no
process is left sleeping on pfmemalloc_wait when kswapd itself goes to
sleep.
However, it isn't necessary to postpone kswapd sleep until the
pfmemalloc_wait queue actually empties. To prevent processes from being
left sleeping, it's actually enough to guarantee that all processes
waiting on pfmemalloc_wait queue have been woken up by the time we put
kswapd to sleep.
This patch therefore fixes this issue by substituting 'wake_up' with
'wake_up_all' and removing 'return false' in the code snippet from
prepare_kswapd_sleep() above. Note that if any process puts itself in
the queue after this waitqueue_active() check, or after the wake up
itself, it means that the process will also wake up kswapd - and since
we are under prepare_to_wait(), the wake up won't be missed. Also we
update the comment prepare_kswapd_sleep() to hopefully more clearly
describe the races it is preventing.
Fixes: 5515061d22f0 ("mm: throttle direct reclaimers if PF_MEMALLOC reserves are low and swap is backed by network storage")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: <stable@vger.kernel.org> [3.6+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-01-08 22:32:40 +00:00
|
|
|
* race between when kswapd checks the watermarks and a process gets
|
|
|
|
* throttled. There is also a potential race if processes get
|
|
|
|
* throttled, kswapd wakes, a large process exits thereby balancing the
|
|
|
|
* zones, which causes kswapd to exit balance_pgdat() before reaching
|
|
|
|
* the wake up checks. If kswapd is going to sleep, no process should
|
|
|
|
* be sleeping on pfmemalloc_wait, so wake them now if necessary. If
|
|
|
|
* the wake up is premature, processes will wake kswapd and get
|
|
|
|
* throttled again. The difference from wake ups in balance_pgdat() is
|
|
|
|
* that here we are under prepare_to_wait().
|
2012-07-31 23:44:35 +00:00
|
|
|
*/
|
mm, vmscan: prevent kswapd livelock due to pfmemalloc-throttled process being killed
Charles Shirron and Paul Cassella from Cray Inc have reported kswapd
stuck in a busy loop with nothing left to balance, but
kswapd_try_to_sleep() failing to sleep. Their analysis found the cause
to be a combination of several factors:
1. A process is waiting in throttle_direct_reclaim() on pgdat->pfmemalloc_wait
2. The process has been killed (by OOM in this case), but has not yet been
scheduled to remove itself from the waitqueue and die.
3. kswapd checks for throttled processes in prepare_kswapd_sleep():
if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
wake_up(&pgdat->pfmemalloc_wait);
return false; // kswapd will not go to sleep
}
However, for a process that was already killed, wake_up() does not remove
the process from the waitqueue, since try_to_wake_up() checks its state
first and returns false when the process is no longer waiting.
4. kswapd is running on the same CPU as the only CPU that the process is
allowed to run on (through cpus_allowed, or possibly single-cpu system).
5. CONFIG_PREEMPT_NONE=y kernel is used. If there's nothing to balance, kswapd
encounters no voluntary preemption points and repeatedly fails
prepare_kswapd_sleep(), blocking the process from running and removing
itself from the waitqueue, which would let kswapd sleep.
So, the source of the problem is that we prevent kswapd from going to
sleep until there are processes waiting on the pfmemalloc_wait queue,
and a process waiting on a queue is guaranteed to be removed from the
queue only when it gets scheduled. This was done to make sure that no
process is left sleeping on pfmemalloc_wait when kswapd itself goes to
sleep.
However, it isn't necessary to postpone kswapd sleep until the
pfmemalloc_wait queue actually empties. To prevent processes from being
left sleeping, it's actually enough to guarantee that all processes
waiting on pfmemalloc_wait queue have been woken up by the time we put
kswapd to sleep.
This patch therefore fixes this issue by substituting 'wake_up' with
'wake_up_all' and removing 'return false' in the code snippet from
prepare_kswapd_sleep() above. Note that if any process puts itself in
the queue after this waitqueue_active() check, or after the wake up
itself, it means that the process will also wake up kswapd - and since
we are under prepare_to_wait(), the wake up won't be missed. Also we
update the comment prepare_kswapd_sleep() to hopefully more clearly
describe the races it is preventing.
Fixes: 5515061d22f0 ("mm: throttle direct reclaimers if PF_MEMALLOC reserves are low and swap is backed by network storage")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: <stable@vger.kernel.org> [3.6+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-01-08 22:32:40 +00:00
|
|
|
if (waitqueue_active(&pgdat->pfmemalloc_wait))
|
|
|
|
wake_up_all(&pgdat->pfmemalloc_wait);
|
2009-12-15 01:58:53 +00:00
|
|
|
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
/* Hopeless node, leave it to direct reclaim */
|
|
|
|
if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
|
|
|
|
return true;
|
|
|
|
|
2020-06-03 22:59:01 +00:00
|
|
|
if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
clear_pgdat_congested(pgdat);
|
|
|
|
return true;
|
2016-07-28 22:45:43 +00:00
|
|
|
}
|
|
|
|
|
mm, vmscan: fix zone balance check in prepare_kswapd_sleep
Patch series "Reduce amount of time kswapd sleeps prematurely", v2.
The series is unusual in that the first patch fixes one problem and
introduces other issues that are noted in the changelog. Patch 2 makes
a minor modification that is worth considering on its own but leaves the
kernel in a state where it behaves badly. It's not until patch 3 that
there is an improvement against baseline.
This was mostly motivated by examining Chris Mason's "simoop" benchmark
which puts the VM under similar pressure to HADOOP. It has been
reported that the benchmark has regressed severely during the last
number of releases. While I cannot reproduce all the same problems
Chris experienced due to hardware limitations, there was a number of
problems on a 2-socket machine with a single disk.
simoop latencies
4.11.0-rc1 4.11.0-rc1
vanilla keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 125765.57 ( 49.08%)
These are latencies. Read/write are threads reading fixed-size random
blocks from a simulated database. The allocation latency is mmaping and
faulting regions of memory. The p50, 95 and p99 reports the worst
latencies for 50% of the samples, 95% and 99% respectively.
For example, the report indicates that while the test was running 99% of
writes completed 99.75% faster. It's worth noting that on a UMA machine
that no difference in performance with simoop was observed so milage
will vary.
It's noted that there is a slight impact to read latencies but it's
mostly due to IO scheduler decisions and offset by the large reduction
in other latencies.
This patch (of 3):
The check in prepare_kswapd_sleep needs to match the one in
balance_pgdat since the latter will return as soon as any one of the
zones in the classzone is above the watermark. This is specially
important for higher order allocations since balance_pgdat will
typically reset the order to zero relying on compaction to create the
higher order pages. Without this patch, prepare_kswapd_sleep fails to
wake up kcompactd since the zone balance check fails.
It was first reported against 4.9.7 that kswapd is failing to wake up
kcompactd due to a mismatch in the zone balance check between
balance_pgdat() and prepare_kswapd_sleep().
balance_pgdat() returns as soon as a single zone satisfies the
allocation but prepare_kswapd_sleep() requires all zones to do +the
same. This causes prepare_kswapd_sleep() to never succeed except in the
order == 0 case and consequently, wakeup_kcompactd() is never called.
For the machine that originally motivated this patch, the state of
compaction from /proc/vmstat looked this way after a day and a half +of
uptime:
compact_migrate_scanned 240496
compact_free_scanned 76238632
compact_isolated 123472
compact_stall 1791
compact_fail 29
compact_success 1762
compact_daemon_wake 0
After applying the patch and about 10 hours of uptime the state looks
like this:
compact_migrate_scanned 59927299
compact_free_scanned 2021075136
compact_isolated 640926
compact_stall 4
compact_fail 2
compact_success 2
compact_daemon_wake 5160
Further notes from Mel that motivated him to pick this patch up and
resend it;
It was observed for the simoop workload (pressures the VM similar to
HADOOP) that kswapd was failing to keep ahead of direct reclaim. The
investigation noted that there was a need to rationalise kswapd
decisions to reclaim with kswapd decisions to sleep. With this patch on
a 2-socket box, there was a 49% reduction in direct reclaim scanning.
However, the impact otherwise is extremely negative. Kswapd reclaim
efficiency dropped from 98% to 76%. simoop has three latency-related
metrics for read, write and allocation (an anonymous mmap and fault).
4.11.0-rc1 4.11.0-rc1
vanilla fixcheck-v2
Amean p50-Read 21670074.18 ( 0.00%) 20464344.18 ( 5.56%)
Amean p95-Read 25456267.64 ( 0.00%) 25721423.64 ( -1.04%)
Amean p99-Read 29369064.73 ( 0.00%) 30174230.76 ( -2.74%)
Amean p50-Write 1390.30 ( 0.00%) 1395.28 ( -0.36%)
Amean p95-Write 412901.57 ( 0.00%) 37737.74 ( 90.86%)
Amean p99-Write 6668722.09 ( 0.00%) 666489.04 ( 90.01%)
Amean p50-Allocation 78714.31 ( 0.00%) 86286.22 ( -9.62%)
Amean p95-Allocation 175533.51 ( 0.00%) 351812.27 (-100.42%)
Amean p99-Allocation 247003.02 ( 0.00%) 6291171.56 (-2447.00%)
Of greater concern is that the patch causes swapping and page writes
from kswapd context rose from 0 pages to 4189753 pages during the hour
the workload ran for. By and large, the patch has very bad behaviour
but easily missed as the impact on a UMA machine is negligible.
This patch is included with the data in case a bisection leads to this
area. This patch is also a pre-requisite for the rest of the series.
Link: http://lkml.kernel.org/r/20170309075657.25121-2-mgorman@techsingularity.net
Signed-off-by: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:38 +00:00
|
|
|
return false;
|
2009-12-15 01:58:53 +00:00
|
|
|
}
|
|
|
|
|
mm: vmscan: limit the number of pages kswapd reclaims at each priority
This series does not fix all the current known problems with reclaim but
it addresses one important swapping bug when there is background IO.
Changelog since V3
- Drop the slab shrink changes in light of Glaubers series and
discussions highlighted that there were a number of potential
problems with the patch. (mel)
- Rebased to 3.10-rc1
Changelog since V2
- Preserve ratio properly for proportional scanning (kamezawa)
Changelog since V1
- Rename ZONE_DIRTY to ZONE_TAIL_LRU_DIRTY (andi)
- Reformat comment in shrink_page_list (andi)
- Clarify some comments (dhillf)
- Rework how the proportional scanning is preserved
- Add PageReclaim check before kswapd starts writeback
- Reset sc.nr_reclaimed on every full zone scan
Kswapd and page reclaim behaviour has been screwy in one way or the
other for a long time. Very broadly speaking it worked in the far past
because machines were limited in memory so it did not have that many
pages to scan and it stalled congestion_wait() frequently to prevent it
going completely nuts. In recent times it has behaved very
unsatisfactorily with some of the problems compounded by the removal of
stall logic and the introduction of transparent hugepage support with
high-order reclaims.
There are many variations of bugs that are rooted in this area. One
example is reports of a large copy operations or backup causing the
machine to grind to a halt or applications pushed to swap. Sometimes in
low memory situations a large percentage of memory suddenly gets
reclaimed. In other cases an application starts and kswapd hits 100%
CPU usage for prolonged periods of time and so on. There is now talk of
introducing features like an extra free kbytes tunable to work around
aspects of the problem instead of trying to deal with it. It's
compounded by the problem that it can be very workload and machine
specific.
This series aims at addressing some of the worst of these problems
without attempting to fundmentally alter how page reclaim works.
Patches 1-2 limits the number of pages kswapd reclaims while still obeying
the anon/file proportion of the LRUs it should be scanning.
Patches 3-4 control how and when kswapd raises its scanning priority and
deletes the scanning restart logic which is tricky to follow.
Patch 5 notes that it is too easy for kswapd to reach priority 0 when
scanning and then reclaim the world. Down with that sort of thing.
Patch 6 notes that kswapd starts writeback based on scanning priority which
is not necessarily related to dirty pages. It will have kswapd
writeback pages if a number of unqueued dirty pages have been
recently encountered at the tail of the LRU.
Patch 7 notes that sometimes kswapd should stall waiting on IO to complete
to reduce LRU churn and the likelihood that it'll reclaim young
clean pages or push applications to swap. It will cause kswapd
to block on IO if it detects that pages being reclaimed under
writeback are recycling through the LRU before the IO completes.
Patchies 8-9 are cosmetic but balance_pgdat() is easier to follow after they
are applied.
This was tested using memcached+memcachetest while some background IO
was in progress as implemented by the parallel IO tests implement in MM
Tests.
memcachetest benchmarks how many operations/second memcached can service
and it is run multiple times. It starts with no background IO and then
re-runs the test with larger amounts of IO in the background to roughly
simulate a large copy in progress. The expectation is that the IO
should have little or no impact on memcachetest which is running
entirely in memory.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Ops memcachetest-0M 22155.00 ( 0.00%) 22180.00 ( 0.11%)
Ops memcachetest-715M 22720.00 ( 0.00%) 22355.00 ( -1.61%)
Ops memcachetest-2385M 3939.00 ( 0.00%) 23450.00 (495.33%)
Ops memcachetest-4055M 3628.00 ( 0.00%) 24341.00 (570.92%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%)
Ops io-duration-2385M 118.00 ( 0.00%) 21.00 ( 82.20%)
Ops io-duration-4055M 162.00 ( 0.00%) 36.00 ( 77.78%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140134.00 ( 0.00%) 18.00 ( 99.99%)
Ops swaptotal-2385M 392438.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 449037.00 ( 0.00%) 27864.00 ( 93.79%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 148031.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 135109.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1529984.00 ( 0.00%) 1530235.00 ( -0.02%)
Ops minorfaults-715M 1794168.00 ( 0.00%) 1613750.00 ( 10.06%)
Ops minorfaults-2385M 1739813.00 ( 0.00%) 1609396.00 ( 7.50%)
Ops minorfaults-4055M 1754460.00 ( 0.00%) 1614810.00 ( 7.96%)
Ops majorfaults-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 185.00 ( 0.00%) 180.00 ( 2.70%)
Ops majorfaults-2385M 24472.00 ( 0.00%) 101.00 ( 99.59%)
Ops majorfaults-4055M 22302.00 ( 0.00%) 229.00 ( 98.97%)
Note how the vanilla kernels performance collapses when there is enough
IO taking place in the background. This drop in performance is part of
what users complain of when they start backups. Note how the swapin and
major fault figures indicate that processes were being pushed to swap
prematurely. With the series applied, there is no noticable performance
drop and while there is still some swap activity, it's tiny.
20 iterations of this test were run in total and averaged. Every 5
iterations, additional IO was generated in the background using dd to
measure how the workload was impacted. The 0M, 715M, 2385M and 4055M
subblock refer to the amount of IO going on in the background at each
iteration. So memcachetest-2385M is reporting how many
transactions/second memcachetest recorded on average over 5 iterations
while there was 2385M of IO going on in the ground. There are six
blocks of information reported here
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
22K/sec to just under 4K/second when there is 2385M of IO going
on in the background. This is one type of performance collapse
users complain about if a large cp or backup starts in the
background
io-duration refers to how long it takes for the background IO to
complete. It's showing that with the patched kernel that the IO
completes faster while not interfering with the memcache
workload
swaptotal is the total amount of swap traffic. With the patched kernel,
the total amount of swapping is much reduced although it is
still not zero.
swapin in this case is an indication as to whether we are swap trashing.
The closer the swapin/swapout ratio is to 1, the worse the
trashing is. Note with the patched kernel that there is no swapin
activity indicating that all the pages swapped were really inactive
unused pages.
minorfaults are just minor faults. An increased number of minor faults
can indicate that page reclaim is unmapping the pages but not
swapping them out before they are faulted back in. With the
patched kernel, there is only a small change in minor faults
majorfaults are just major faults in the target workload and a high
number can indicate that a workload is being prematurely
swapped. With the patched kernel, major faults are much reduced. As
there are no swapin's recorded so it's not being swapped. The likely
explanation is that that libraries or configuration files used by
the workload during startup get paged out by the background IO.
Overall with the series applied, there is no noticable performance drop
due to background IO and while there is still some swap activity, it's
tiny and the lack of swapins imply that the swapped pages were inactive
and unused.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Page Ins 1234608 101892
Page Outs 12446272 11810468
Swap Ins 283406 0
Swap Outs 698469 27882
Direct pages scanned 0 136480
Kswapd pages scanned 6266537 5369364
Kswapd pages reclaimed 1088989 930832
Direct pages reclaimed 0 120901
Kswapd efficiency 17% 17%
Kswapd velocity 5398.371 4635.115
Direct efficiency 100% 88%
Direct velocity 0.000 117.817
Percentage direct scans 0% 2%
Page writes by reclaim 1655843 4009929
Page writes file 957374 3982047
Page writes anon 698469 27882
Page reclaim immediate 5245 1745
Page rescued immediate 0 0
Slabs scanned 33664 25216
Direct inode steals 0 0
Kswapd inode steals 19409 778
Kswapd skipped wait 0 0
THP fault alloc 35 30
THP collapse alloc 472 401
THP splits 27 22
THP fault fallback 0 0
THP collapse fail 0 1
Compaction stalls 0 4
Compaction success 0 0
Compaction failures 0 4
Page migrate success 0 0
Page migrate failure 0 0
Compaction pages isolated 0 0
Compaction migrate scanned 0 0
Compaction free scanned 0 0
Compaction cost 0 0
NUMA PTE updates 0 0
NUMA hint faults 0 0
NUMA hint local faults 0 0
NUMA pages migrated 0 0
AutoNUMA cost 0 0
Unfortunately, note that there is a small amount of direct reclaim due to
kswapd no longer reclaiming the world. ftrace indicates that the direct
reclaim stalls are mostly harmless with the vast bulk of the stalls
incurred by dd
23 tclsh-3367
38 memcachetest-13733
49 memcachetest-12443
57 tee-3368
1541 dd-13826
1981 dd-12539
A consequence of the direct reclaim for dd is that the processes for the
IO workload may show a higher system CPU usage. There is also a risk that
kswapd not reclaiming the world may mean that it stays awake balancing
zones, does not stall on the appropriate events and continually scans
pages it cannot reclaim consuming CPU. This will be visible as continued
high CPU usage but in my own tests I only saw a single spike lasting less
than a second and I did not observe any problems related to reclaim while
running the series on my desktop.
This patch:
The number of pages kswapd can reclaim is bound by the number of pages it
scans which is related to the size of the zone and the scanning priority.
In many cases the priority remains low because it's reset every
SWAP_CLUSTER_MAX reclaimed pages but in the event kswapd scans a large
number of pages it cannot reclaim, it will raise the priority and
potentially discard a large percentage of the zone as sc->nr_to_reclaim is
ULONG_MAX. The user-visible effect is a reclaim "spike" where a large
percentage of memory is suddenly freed. It would be bad enough if this
was just unused memory but because of how anon/file pages are balanced it
is possible that applications get pushed to swap unnecessarily.
This patch limits the number of pages kswapd will reclaim to the high
watermark. Reclaim will still overshoot due to it not being a hard limit
as shrink_lruvec() will ignore the sc.nr_to_reclaim at DEF_PRIORITY but it
prevents kswapd reclaiming the world at higher priorities. The number of
pages it reclaims is not adjusted for high-order allocations as kswapd
will reclaim excessively if it is to balance zones for high-order
allocations.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Tested-by: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:42 +00:00
|
|
|
/*
|
2016-07-28 22:45:43 +00:00
|
|
|
* kswapd shrinks a node of pages that are at or below the highest usable
|
|
|
|
* zone that is currently unbalanced.
|
2013-07-03 22:01:45 +00:00
|
|
|
*
|
|
|
|
* Returns true if kswapd scanned at least the requested number of pages to
|
2013-07-03 22:01:51 +00:00
|
|
|
* reclaim or if the lack of progress was due to pages under writeback.
|
|
|
|
* This is used to determine if the scanning priority needs to be raised.
|
mm: vmscan: limit the number of pages kswapd reclaims at each priority
This series does not fix all the current known problems with reclaim but
it addresses one important swapping bug when there is background IO.
Changelog since V3
- Drop the slab shrink changes in light of Glaubers series and
discussions highlighted that there were a number of potential
problems with the patch. (mel)
- Rebased to 3.10-rc1
Changelog since V2
- Preserve ratio properly for proportional scanning (kamezawa)
Changelog since V1
- Rename ZONE_DIRTY to ZONE_TAIL_LRU_DIRTY (andi)
- Reformat comment in shrink_page_list (andi)
- Clarify some comments (dhillf)
- Rework how the proportional scanning is preserved
- Add PageReclaim check before kswapd starts writeback
- Reset sc.nr_reclaimed on every full zone scan
Kswapd and page reclaim behaviour has been screwy in one way or the
other for a long time. Very broadly speaking it worked in the far past
because machines were limited in memory so it did not have that many
pages to scan and it stalled congestion_wait() frequently to prevent it
going completely nuts. In recent times it has behaved very
unsatisfactorily with some of the problems compounded by the removal of
stall logic and the introduction of transparent hugepage support with
high-order reclaims.
There are many variations of bugs that are rooted in this area. One
example is reports of a large copy operations or backup causing the
machine to grind to a halt or applications pushed to swap. Sometimes in
low memory situations a large percentage of memory suddenly gets
reclaimed. In other cases an application starts and kswapd hits 100%
CPU usage for prolonged periods of time and so on. There is now talk of
introducing features like an extra free kbytes tunable to work around
aspects of the problem instead of trying to deal with it. It's
compounded by the problem that it can be very workload and machine
specific.
This series aims at addressing some of the worst of these problems
without attempting to fundmentally alter how page reclaim works.
Patches 1-2 limits the number of pages kswapd reclaims while still obeying
the anon/file proportion of the LRUs it should be scanning.
Patches 3-4 control how and when kswapd raises its scanning priority and
deletes the scanning restart logic which is tricky to follow.
Patch 5 notes that it is too easy for kswapd to reach priority 0 when
scanning and then reclaim the world. Down with that sort of thing.
Patch 6 notes that kswapd starts writeback based on scanning priority which
is not necessarily related to dirty pages. It will have kswapd
writeback pages if a number of unqueued dirty pages have been
recently encountered at the tail of the LRU.
Patch 7 notes that sometimes kswapd should stall waiting on IO to complete
to reduce LRU churn and the likelihood that it'll reclaim young
clean pages or push applications to swap. It will cause kswapd
to block on IO if it detects that pages being reclaimed under
writeback are recycling through the LRU before the IO completes.
Patchies 8-9 are cosmetic but balance_pgdat() is easier to follow after they
are applied.
This was tested using memcached+memcachetest while some background IO
was in progress as implemented by the parallel IO tests implement in MM
Tests.
memcachetest benchmarks how many operations/second memcached can service
and it is run multiple times. It starts with no background IO and then
re-runs the test with larger amounts of IO in the background to roughly
simulate a large copy in progress. The expectation is that the IO
should have little or no impact on memcachetest which is running
entirely in memory.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Ops memcachetest-0M 22155.00 ( 0.00%) 22180.00 ( 0.11%)
Ops memcachetest-715M 22720.00 ( 0.00%) 22355.00 ( -1.61%)
Ops memcachetest-2385M 3939.00 ( 0.00%) 23450.00 (495.33%)
Ops memcachetest-4055M 3628.00 ( 0.00%) 24341.00 (570.92%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%)
Ops io-duration-2385M 118.00 ( 0.00%) 21.00 ( 82.20%)
Ops io-duration-4055M 162.00 ( 0.00%) 36.00 ( 77.78%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140134.00 ( 0.00%) 18.00 ( 99.99%)
Ops swaptotal-2385M 392438.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 449037.00 ( 0.00%) 27864.00 ( 93.79%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 148031.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 135109.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1529984.00 ( 0.00%) 1530235.00 ( -0.02%)
Ops minorfaults-715M 1794168.00 ( 0.00%) 1613750.00 ( 10.06%)
Ops minorfaults-2385M 1739813.00 ( 0.00%) 1609396.00 ( 7.50%)
Ops minorfaults-4055M 1754460.00 ( 0.00%) 1614810.00 ( 7.96%)
Ops majorfaults-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 185.00 ( 0.00%) 180.00 ( 2.70%)
Ops majorfaults-2385M 24472.00 ( 0.00%) 101.00 ( 99.59%)
Ops majorfaults-4055M 22302.00 ( 0.00%) 229.00 ( 98.97%)
Note how the vanilla kernels performance collapses when there is enough
IO taking place in the background. This drop in performance is part of
what users complain of when they start backups. Note how the swapin and
major fault figures indicate that processes were being pushed to swap
prematurely. With the series applied, there is no noticable performance
drop and while there is still some swap activity, it's tiny.
20 iterations of this test were run in total and averaged. Every 5
iterations, additional IO was generated in the background using dd to
measure how the workload was impacted. The 0M, 715M, 2385M and 4055M
subblock refer to the amount of IO going on in the background at each
iteration. So memcachetest-2385M is reporting how many
transactions/second memcachetest recorded on average over 5 iterations
while there was 2385M of IO going on in the ground. There are six
blocks of information reported here
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
22K/sec to just under 4K/second when there is 2385M of IO going
on in the background. This is one type of performance collapse
users complain about if a large cp or backup starts in the
background
io-duration refers to how long it takes for the background IO to
complete. It's showing that with the patched kernel that the IO
completes faster while not interfering with the memcache
workload
swaptotal is the total amount of swap traffic. With the patched kernel,
the total amount of swapping is much reduced although it is
still not zero.
swapin in this case is an indication as to whether we are swap trashing.
The closer the swapin/swapout ratio is to 1, the worse the
trashing is. Note with the patched kernel that there is no swapin
activity indicating that all the pages swapped were really inactive
unused pages.
minorfaults are just minor faults. An increased number of minor faults
can indicate that page reclaim is unmapping the pages but not
swapping them out before they are faulted back in. With the
patched kernel, there is only a small change in minor faults
majorfaults are just major faults in the target workload and a high
number can indicate that a workload is being prematurely
swapped. With the patched kernel, major faults are much reduced. As
there are no swapin's recorded so it's not being swapped. The likely
explanation is that that libraries or configuration files used by
the workload during startup get paged out by the background IO.
Overall with the series applied, there is no noticable performance drop
due to background IO and while there is still some swap activity, it's
tiny and the lack of swapins imply that the swapped pages were inactive
and unused.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Page Ins 1234608 101892
Page Outs 12446272 11810468
Swap Ins 283406 0
Swap Outs 698469 27882
Direct pages scanned 0 136480
Kswapd pages scanned 6266537 5369364
Kswapd pages reclaimed 1088989 930832
Direct pages reclaimed 0 120901
Kswapd efficiency 17% 17%
Kswapd velocity 5398.371 4635.115
Direct efficiency 100% 88%
Direct velocity 0.000 117.817
Percentage direct scans 0% 2%
Page writes by reclaim 1655843 4009929
Page writes file 957374 3982047
Page writes anon 698469 27882
Page reclaim immediate 5245 1745
Page rescued immediate 0 0
Slabs scanned 33664 25216
Direct inode steals 0 0
Kswapd inode steals 19409 778
Kswapd skipped wait 0 0
THP fault alloc 35 30
THP collapse alloc 472 401
THP splits 27 22
THP fault fallback 0 0
THP collapse fail 0 1
Compaction stalls 0 4
Compaction success 0 0
Compaction failures 0 4
Page migrate success 0 0
Page migrate failure 0 0
Compaction pages isolated 0 0
Compaction migrate scanned 0 0
Compaction free scanned 0 0
Compaction cost 0 0
NUMA PTE updates 0 0
NUMA hint faults 0 0
NUMA hint local faults 0 0
NUMA pages migrated 0 0
AutoNUMA cost 0 0
Unfortunately, note that there is a small amount of direct reclaim due to
kswapd no longer reclaiming the world. ftrace indicates that the direct
reclaim stalls are mostly harmless with the vast bulk of the stalls
incurred by dd
23 tclsh-3367
38 memcachetest-13733
49 memcachetest-12443
57 tee-3368
1541 dd-13826
1981 dd-12539
A consequence of the direct reclaim for dd is that the processes for the
IO workload may show a higher system CPU usage. There is also a risk that
kswapd not reclaiming the world may mean that it stays awake balancing
zones, does not stall on the appropriate events and continually scans
pages it cannot reclaim consuming CPU. This will be visible as continued
high CPU usage but in my own tests I only saw a single spike lasting less
than a second and I did not observe any problems related to reclaim while
running the series on my desktop.
This patch:
The number of pages kswapd can reclaim is bound by the number of pages it
scans which is related to the size of the zone and the scanning priority.
In many cases the priority remains low because it's reset every
SWAP_CLUSTER_MAX reclaimed pages but in the event kswapd scans a large
number of pages it cannot reclaim, it will raise the priority and
potentially discard a large percentage of the zone as sc->nr_to_reclaim is
ULONG_MAX. The user-visible effect is a reclaim "spike" where a large
percentage of memory is suddenly freed. It would be bad enough if this
was just unused memory but because of how anon/file pages are balanced it
is possible that applications get pushed to swap unnecessarily.
This patch limits the number of pages kswapd will reclaim to the high
watermark. Reclaim will still overshoot due to it not being a hard limit
as shrink_lruvec() will ignore the sc.nr_to_reclaim at DEF_PRIORITY but it
prevents kswapd reclaiming the world at higher priorities. The number of
pages it reclaims is not adjusted for high-order allocations as kswapd
will reclaim excessively if it is to balance zones for high-order
allocations.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Tested-by: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:42 +00:00
|
|
|
*/
|
2016-07-28 22:45:43 +00:00
|
|
|
static bool kswapd_shrink_node(pg_data_t *pgdat,
|
mm, kswapd: replace kswapd compaction with waking up kcompactd
Similarly to direct reclaim/compaction, kswapd attempts to combine
reclaim and compaction to attempt making memory allocation of given
order available.
The details differ from direct reclaim e.g. in having high watermark as
a goal. The code involved in kswapd's reclaim/compaction decisions has
evolved to be quite complex.
Testing reveals that it doesn't actually work in at least one scenario,
and closer inspection suggests that it could be greatly simplified
without compromising on the goal (make high-order page available) or
efficiency (don't reclaim too much). The simplification relieas of
doing all compaction in kcompactd, which is simply woken up when high
watermarks are reached by kswapd's reclaim.
The scenario where kswapd compaction doesn't work was found with mmtests
test stress-highalloc configured to attempt order-9 allocations without
direct reclaim, just waking up kswapd. There was no compaction attempt
from kswapd during the whole test. Some added instrumentation shows
what happens:
- balance_pgdat() sets end_zone to Normal, as it's not balanced
- reclaim is attempted on DMA zone, which sets nr_attempted to 99, but
it cannot reclaim anything, so sc.nr_reclaimed is 0
- for zones DMA32 and Normal, kswapd_shrink_zone uses testorder=0, so
it merely checks if high watermarks were reached for base pages.
This is true, so no reclaim is attempted. For DMA, testorder=0
wasn't used, as compaction_suitable() returned COMPACT_SKIPPED
- even though the pgdat_needs_compaction flag wasn't set to false, no
compaction happens due to the condition sc.nr_reclaimed >
nr_attempted being false (as 0 < 99)
- priority-- due to nr_reclaimed being 0, repeat until priority reaches
0 pgdat_balanced() is false as only the small zone DMA appears
balanced (curiously in that check, watermark appears OK and
compaction_suitable() returns COMPACT_PARTIAL, because a lower
classzone_idx is used there)
Now, even if it was decided that reclaim shouldn't be attempted on the
DMA zone, the scenario would be the same, as (sc.nr_reclaimed=0 >
nr_attempted=0) is also false. The condition really should use >= as
the comment suggests. Then there is a mismatch in the check for setting
pgdat_needs_compaction to false using low watermark, while the rest uses
high watermark, and who knows what other subtlety. Hopefully this
demonstrates that this is unsustainable.
Luckily we can simplify this a lot. The reclaim/compaction decisions
make sense for direct reclaim scenario, but in kswapd, our primary goal
is to reach high watermark in order-0 pages. Afterwards we can attempt
compaction just once. Unlike direct reclaim, we don't reclaim extra
pages (over the high watermark), the current code already disallows it
for good reasons.
After this patch, we simply wake up kcompactd to process the pgdat,
after we have either succeeded or failed to reach the high watermarks in
kswapd, which goes to sleep. We pass kswapd's order and classzone_idx,
so kcompactd can apply the same criteria to determine which zones are
worth compacting. Note that we use the classzone_idx from
wakeup_kswapd(), not balanced_classzone_idx which can include higher
zones that kswapd tried to balance too, but didn't consider them in
pgdat_balanced().
Since kswapd now cannot create high-order pages itself, we need to
adjust how it determines the zones to be balanced. The key element here
is adding a "highorder" parameter to zone_balanced, which, when set to
false, makes it consider only order-0 watermark instead of the desired
higher order (this was done previously by kswapd_shrink_zone(), but not
elsewhere). This false is passed for example in pgdat_balanced().
Importantly, wakeup_kswapd() uses true to make sure kswapd and thus
kcompactd are woken up for a high-order allocation failure.
The last thing is to decide what to do with pageblock_skip bitmap
handling. Compaction maintains a pageblock_skip bitmap to record
pageblocks where isolation recently failed. This bitmap can be reset by
three ways:
1) direct compaction is restarting after going through the full deferred cycle
2) kswapd goes to sleep, and some other direct compaction has previously
finished scanning the whole zone and set zone->compact_blockskip_flush.
Note that a successful direct compaction clears this flag.
3) compaction was invoked manually via trigger in /proc
The case 2) is somewhat fuzzy to begin with, but after introducing
kcompactd we should update it. The check for direct compaction in 1),
and to set the flush flag in 2) use current_is_kswapd(), which doesn't
work for kcompactd. Thus, this patch adds bool direct_compaction to
compact_control to use in 2). For the case 1) we remove the check
completely - unlike the former kswapd compaction, kcompactd does use the
deferred compaction functionality, so flushing tied to restarting from
deferred compaction makes sense here.
Note that when kswapd goes to sleep, kcompactd is woken up, so it will
see the flushed pageblock_skip bits. This is different from when the
former kswapd compaction observed the bits and I believe it makes more
sense. Kcompactd can afford to be more thorough than a direct
compaction trying to limit allocation latency, or kswapd whose primary
goal is to reclaim.
For testing, I used stress-highalloc configured to do order-9
allocations with GFP_NOWAIT|__GFP_HIGH|__GFP_COMP, so they relied just
on kswapd/kcompactd reclaim/compaction (the interfering kernel builds in
phases 1 and 2 work as usual):
stress-highalloc
4.5-rc1+before 4.5-rc1+after
-nodirect -nodirect
Success 1 Min 1.00 ( 0.00%) 5.00 (-66.67%)
Success 1 Mean 1.40 ( 0.00%) 6.20 (-55.00%)
Success 1 Max 2.00 ( 0.00%) 7.00 (-16.67%)
Success 2 Min 1.00 ( 0.00%) 5.00 (-66.67%)
Success 2 Mean 1.80 ( 0.00%) 6.40 (-52.38%)
Success 2 Max 3.00 ( 0.00%) 7.00 (-16.67%)
Success 3 Min 34.00 ( 0.00%) 62.00 ( 1.59%)
Success 3 Mean 41.80 ( 0.00%) 63.80 ( 1.24%)
Success 3 Max 53.00 ( 0.00%) 65.00 ( 2.99%)
User 3166.67 3181.09
System 1153.37 1158.25
Elapsed 1768.53 1799.37
4.5-rc1+before 4.5-rc1+after
-nodirect -nodirect
Direct pages scanned 32938 32797
Kswapd pages scanned 2183166 2202613
Kswapd pages reclaimed 2152359 2143524
Direct pages reclaimed 32735 32545
Percentage direct scans 1% 1%
THP fault alloc 579 612
THP collapse alloc 304 316
THP splits 0 0
THP fault fallback 793 778
THP collapse fail 11 16
Compaction stalls 1013 1007
Compaction success 92 67
Compaction failures 920 939
Page migrate success 238457 721374
Page migrate failure 23021 23469
Compaction pages isolated 504695 1479924
Compaction migrate scanned 661390 8812554
Compaction free scanned 13476658 84327916
Compaction cost 262 838
After this patch we see improvements in allocation success rate
(especially for phase 3) along with increased compaction activity. The
compaction stalls (direct compaction) in the interfering kernel builds
(probably THP's) also decreased somewhat thanks to kcompactd activity,
yet THP alloc successes improved a bit.
Note that elapsed and user time isn't so useful for this benchmark,
because of the background interference being unpredictable. It's just
to quickly spot some major unexpected differences. System time is
somewhat more useful and that didn't increase.
Also (after adjusting mmtests' ftrace monitor):
Time kswapd awake 2547781 2269241
Time kcompactd awake 0 119253
Time direct compacting 939937 557649
Time kswapd compacting 0 0
Time kcompactd compacting 0 119099
The decrease of overal time spent compacting appears to not match the
increased compaction stats. I suspect the tasks get rescheduled and
since the ftrace monitor doesn't see that, the reported time is wall
time, not CPU time. But arguably direct compactors care about overall
latency anyway, whether busy compacting or waiting for CPU doesn't
matter. And that latency seems to almost halved.
It's also interesting how much time kswapd spent awake just going
through all the priorities and failing to even try compacting, over and
over.
We can also configure stress-highalloc to perform both direct
reclaim/compaction and wakeup kswapd/kcompactd, by using
GFP_KERNEL|__GFP_HIGH|__GFP_COMP:
stress-highalloc
4.5-rc1+before 4.5-rc1+after
-direct -direct
Success 1 Min 4.00 ( 0.00%) 9.00 (-50.00%)
Success 1 Mean 8.00 ( 0.00%) 10.00 (-19.05%)
Success 1 Max 12.00 ( 0.00%) 11.00 ( 15.38%)
Success 2 Min 4.00 ( 0.00%) 9.00 (-50.00%)
Success 2 Mean 8.20 ( 0.00%) 10.00 (-16.28%)
Success 2 Max 13.00 ( 0.00%) 11.00 ( 8.33%)
Success 3 Min 75.00 ( 0.00%) 74.00 ( 1.33%)
Success 3 Mean 75.60 ( 0.00%) 75.20 ( 0.53%)
Success 3 Max 77.00 ( 0.00%) 76.00 ( 0.00%)
User 3344.73 3246.04
System 1194.24 1172.29
Elapsed 1838.04 1836.76
4.5-rc1+before 4.5-rc1+after
-direct -direct
Direct pages scanned 125146 120966
Kswapd pages scanned 2119757 2135012
Kswapd pages reclaimed 2073183 2108388
Direct pages reclaimed 124909 120577
Percentage direct scans 5% 5%
THP fault alloc 599 652
THP collapse alloc 323 354
THP splits 0 0
THP fault fallback 806 793
THP collapse fail 17 16
Compaction stalls 2457 2025
Compaction success 906 518
Compaction failures 1551 1507
Page migrate success 2031423 2360608
Page migrate failure 32845 40852
Compaction pages isolated 4129761 4802025
Compaction migrate scanned 11996712 21750613
Compaction free scanned 214970969 344372001
Compaction cost 2271 2694
In this scenario, this patch doesn't change the overall success rate as
direct compaction already tries all it can. There's however significant
reduction in direct compaction stalls (that is, the number of
allocations that went into direct compaction). The number of successes
(i.e. direct compaction stalls that ended up with successful
allocation) is reduced by the same number. This means the offload to
kcompactd is working as expected, and direct compaction is reduced
either due to detecting contention, or compaction deferred by kcompactd.
In the previous version of this patchset there was some apparent
reduction of success rate, but the changes in this version (such as
using sync compaction only), new baseline kernel, and/or averaging
results from 5 executions (my bet), made this go away.
Ftrace-based stats seem to roughly agree:
Time kswapd awake 2532984 2326824
Time kcompactd awake 0 257916
Time direct compacting 864839 735130
Time kswapd compacting 0 0
Time kcompactd compacting 0 257585
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-17 21:18:15 +00:00
|
|
|
struct scan_control *sc)
|
mm: vmscan: limit the number of pages kswapd reclaims at each priority
This series does not fix all the current known problems with reclaim but
it addresses one important swapping bug when there is background IO.
Changelog since V3
- Drop the slab shrink changes in light of Glaubers series and
discussions highlighted that there were a number of potential
problems with the patch. (mel)
- Rebased to 3.10-rc1
Changelog since V2
- Preserve ratio properly for proportional scanning (kamezawa)
Changelog since V1
- Rename ZONE_DIRTY to ZONE_TAIL_LRU_DIRTY (andi)
- Reformat comment in shrink_page_list (andi)
- Clarify some comments (dhillf)
- Rework how the proportional scanning is preserved
- Add PageReclaim check before kswapd starts writeback
- Reset sc.nr_reclaimed on every full zone scan
Kswapd and page reclaim behaviour has been screwy in one way or the
other for a long time. Very broadly speaking it worked in the far past
because machines were limited in memory so it did not have that many
pages to scan and it stalled congestion_wait() frequently to prevent it
going completely nuts. In recent times it has behaved very
unsatisfactorily with some of the problems compounded by the removal of
stall logic and the introduction of transparent hugepage support with
high-order reclaims.
There are many variations of bugs that are rooted in this area. One
example is reports of a large copy operations or backup causing the
machine to grind to a halt or applications pushed to swap. Sometimes in
low memory situations a large percentage of memory suddenly gets
reclaimed. In other cases an application starts and kswapd hits 100%
CPU usage for prolonged periods of time and so on. There is now talk of
introducing features like an extra free kbytes tunable to work around
aspects of the problem instead of trying to deal with it. It's
compounded by the problem that it can be very workload and machine
specific.
This series aims at addressing some of the worst of these problems
without attempting to fundmentally alter how page reclaim works.
Patches 1-2 limits the number of pages kswapd reclaims while still obeying
the anon/file proportion of the LRUs it should be scanning.
Patches 3-4 control how and when kswapd raises its scanning priority and
deletes the scanning restart logic which is tricky to follow.
Patch 5 notes that it is too easy for kswapd to reach priority 0 when
scanning and then reclaim the world. Down with that sort of thing.
Patch 6 notes that kswapd starts writeback based on scanning priority which
is not necessarily related to dirty pages. It will have kswapd
writeback pages if a number of unqueued dirty pages have been
recently encountered at the tail of the LRU.
Patch 7 notes that sometimes kswapd should stall waiting on IO to complete
to reduce LRU churn and the likelihood that it'll reclaim young
clean pages or push applications to swap. It will cause kswapd
to block on IO if it detects that pages being reclaimed under
writeback are recycling through the LRU before the IO completes.
Patchies 8-9 are cosmetic but balance_pgdat() is easier to follow after they
are applied.
This was tested using memcached+memcachetest while some background IO
was in progress as implemented by the parallel IO tests implement in MM
Tests.
memcachetest benchmarks how many operations/second memcached can service
and it is run multiple times. It starts with no background IO and then
re-runs the test with larger amounts of IO in the background to roughly
simulate a large copy in progress. The expectation is that the IO
should have little or no impact on memcachetest which is running
entirely in memory.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Ops memcachetest-0M 22155.00 ( 0.00%) 22180.00 ( 0.11%)
Ops memcachetest-715M 22720.00 ( 0.00%) 22355.00 ( -1.61%)
Ops memcachetest-2385M 3939.00 ( 0.00%) 23450.00 (495.33%)
Ops memcachetest-4055M 3628.00 ( 0.00%) 24341.00 (570.92%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%)
Ops io-duration-2385M 118.00 ( 0.00%) 21.00 ( 82.20%)
Ops io-duration-4055M 162.00 ( 0.00%) 36.00 ( 77.78%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140134.00 ( 0.00%) 18.00 ( 99.99%)
Ops swaptotal-2385M 392438.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 449037.00 ( 0.00%) 27864.00 ( 93.79%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 148031.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 135109.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1529984.00 ( 0.00%) 1530235.00 ( -0.02%)
Ops minorfaults-715M 1794168.00 ( 0.00%) 1613750.00 ( 10.06%)
Ops minorfaults-2385M 1739813.00 ( 0.00%) 1609396.00 ( 7.50%)
Ops minorfaults-4055M 1754460.00 ( 0.00%) 1614810.00 ( 7.96%)
Ops majorfaults-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 185.00 ( 0.00%) 180.00 ( 2.70%)
Ops majorfaults-2385M 24472.00 ( 0.00%) 101.00 ( 99.59%)
Ops majorfaults-4055M 22302.00 ( 0.00%) 229.00 ( 98.97%)
Note how the vanilla kernels performance collapses when there is enough
IO taking place in the background. This drop in performance is part of
what users complain of when they start backups. Note how the swapin and
major fault figures indicate that processes were being pushed to swap
prematurely. With the series applied, there is no noticable performance
drop and while there is still some swap activity, it's tiny.
20 iterations of this test were run in total and averaged. Every 5
iterations, additional IO was generated in the background using dd to
measure how the workload was impacted. The 0M, 715M, 2385M and 4055M
subblock refer to the amount of IO going on in the background at each
iteration. So memcachetest-2385M is reporting how many
transactions/second memcachetest recorded on average over 5 iterations
while there was 2385M of IO going on in the ground. There are six
blocks of information reported here
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
22K/sec to just under 4K/second when there is 2385M of IO going
on in the background. This is one type of performance collapse
users complain about if a large cp or backup starts in the
background
io-duration refers to how long it takes for the background IO to
complete. It's showing that with the patched kernel that the IO
completes faster while not interfering with the memcache
workload
swaptotal is the total amount of swap traffic. With the patched kernel,
the total amount of swapping is much reduced although it is
still not zero.
swapin in this case is an indication as to whether we are swap trashing.
The closer the swapin/swapout ratio is to 1, the worse the
trashing is. Note with the patched kernel that there is no swapin
activity indicating that all the pages swapped were really inactive
unused pages.
minorfaults are just minor faults. An increased number of minor faults
can indicate that page reclaim is unmapping the pages but not
swapping them out before they are faulted back in. With the
patched kernel, there is only a small change in minor faults
majorfaults are just major faults in the target workload and a high
number can indicate that a workload is being prematurely
swapped. With the patched kernel, major faults are much reduced. As
there are no swapin's recorded so it's not being swapped. The likely
explanation is that that libraries or configuration files used by
the workload during startup get paged out by the background IO.
Overall with the series applied, there is no noticable performance drop
due to background IO and while there is still some swap activity, it's
tiny and the lack of swapins imply that the swapped pages were inactive
and unused.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Page Ins 1234608 101892
Page Outs 12446272 11810468
Swap Ins 283406 0
Swap Outs 698469 27882
Direct pages scanned 0 136480
Kswapd pages scanned 6266537 5369364
Kswapd pages reclaimed 1088989 930832
Direct pages reclaimed 0 120901
Kswapd efficiency 17% 17%
Kswapd velocity 5398.371 4635.115
Direct efficiency 100% 88%
Direct velocity 0.000 117.817
Percentage direct scans 0% 2%
Page writes by reclaim 1655843 4009929
Page writes file 957374 3982047
Page writes anon 698469 27882
Page reclaim immediate 5245 1745
Page rescued immediate 0 0
Slabs scanned 33664 25216
Direct inode steals 0 0
Kswapd inode steals 19409 778
Kswapd skipped wait 0 0
THP fault alloc 35 30
THP collapse alloc 472 401
THP splits 27 22
THP fault fallback 0 0
THP collapse fail 0 1
Compaction stalls 0 4
Compaction success 0 0
Compaction failures 0 4
Page migrate success 0 0
Page migrate failure 0 0
Compaction pages isolated 0 0
Compaction migrate scanned 0 0
Compaction free scanned 0 0
Compaction cost 0 0
NUMA PTE updates 0 0
NUMA hint faults 0 0
NUMA hint local faults 0 0
NUMA pages migrated 0 0
AutoNUMA cost 0 0
Unfortunately, note that there is a small amount of direct reclaim due to
kswapd no longer reclaiming the world. ftrace indicates that the direct
reclaim stalls are mostly harmless with the vast bulk of the stalls
incurred by dd
23 tclsh-3367
38 memcachetest-13733
49 memcachetest-12443
57 tee-3368
1541 dd-13826
1981 dd-12539
A consequence of the direct reclaim for dd is that the processes for the
IO workload may show a higher system CPU usage. There is also a risk that
kswapd not reclaiming the world may mean that it stays awake balancing
zones, does not stall on the appropriate events and continually scans
pages it cannot reclaim consuming CPU. This will be visible as continued
high CPU usage but in my own tests I only saw a single spike lasting less
than a second and I did not observe any problems related to reclaim while
running the series on my desktop.
This patch:
The number of pages kswapd can reclaim is bound by the number of pages it
scans which is related to the size of the zone and the scanning priority.
In many cases the priority remains low because it's reset every
SWAP_CLUSTER_MAX reclaimed pages but in the event kswapd scans a large
number of pages it cannot reclaim, it will raise the priority and
potentially discard a large percentage of the zone as sc->nr_to_reclaim is
ULONG_MAX. The user-visible effect is a reclaim "spike" where a large
percentage of memory is suddenly freed. It would be bad enough if this
was just unused memory but because of how anon/file pages are balanced it
is possible that applications get pushed to swap unnecessarily.
This patch limits the number of pages kswapd will reclaim to the high
watermark. Reclaim will still overshoot due to it not being a hard limit
as shrink_lruvec() will ignore the sc.nr_to_reclaim at DEF_PRIORITY but it
prevents kswapd reclaiming the world at higher priorities. The number of
pages it reclaims is not adjusted for high-order allocations as kswapd
will reclaim excessively if it is to balance zones for high-order
allocations.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Tested-by: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:42 +00:00
|
|
|
{
|
2016-07-28 22:45:43 +00:00
|
|
|
struct zone *zone;
|
|
|
|
int z;
|
mm: vmscan: limit the number of pages kswapd reclaims at each priority
This series does not fix all the current known problems with reclaim but
it addresses one important swapping bug when there is background IO.
Changelog since V3
- Drop the slab shrink changes in light of Glaubers series and
discussions highlighted that there were a number of potential
problems with the patch. (mel)
- Rebased to 3.10-rc1
Changelog since V2
- Preserve ratio properly for proportional scanning (kamezawa)
Changelog since V1
- Rename ZONE_DIRTY to ZONE_TAIL_LRU_DIRTY (andi)
- Reformat comment in shrink_page_list (andi)
- Clarify some comments (dhillf)
- Rework how the proportional scanning is preserved
- Add PageReclaim check before kswapd starts writeback
- Reset sc.nr_reclaimed on every full zone scan
Kswapd and page reclaim behaviour has been screwy in one way or the
other for a long time. Very broadly speaking it worked in the far past
because machines were limited in memory so it did not have that many
pages to scan and it stalled congestion_wait() frequently to prevent it
going completely nuts. In recent times it has behaved very
unsatisfactorily with some of the problems compounded by the removal of
stall logic and the introduction of transparent hugepage support with
high-order reclaims.
There are many variations of bugs that are rooted in this area. One
example is reports of a large copy operations or backup causing the
machine to grind to a halt or applications pushed to swap. Sometimes in
low memory situations a large percentage of memory suddenly gets
reclaimed. In other cases an application starts and kswapd hits 100%
CPU usage for prolonged periods of time and so on. There is now talk of
introducing features like an extra free kbytes tunable to work around
aspects of the problem instead of trying to deal with it. It's
compounded by the problem that it can be very workload and machine
specific.
This series aims at addressing some of the worst of these problems
without attempting to fundmentally alter how page reclaim works.
Patches 1-2 limits the number of pages kswapd reclaims while still obeying
the anon/file proportion of the LRUs it should be scanning.
Patches 3-4 control how and when kswapd raises its scanning priority and
deletes the scanning restart logic which is tricky to follow.
Patch 5 notes that it is too easy for kswapd to reach priority 0 when
scanning and then reclaim the world. Down with that sort of thing.
Patch 6 notes that kswapd starts writeback based on scanning priority which
is not necessarily related to dirty pages. It will have kswapd
writeback pages if a number of unqueued dirty pages have been
recently encountered at the tail of the LRU.
Patch 7 notes that sometimes kswapd should stall waiting on IO to complete
to reduce LRU churn and the likelihood that it'll reclaim young
clean pages or push applications to swap. It will cause kswapd
to block on IO if it detects that pages being reclaimed under
writeback are recycling through the LRU before the IO completes.
Patchies 8-9 are cosmetic but balance_pgdat() is easier to follow after they
are applied.
This was tested using memcached+memcachetest while some background IO
was in progress as implemented by the parallel IO tests implement in MM
Tests.
memcachetest benchmarks how many operations/second memcached can service
and it is run multiple times. It starts with no background IO and then
re-runs the test with larger amounts of IO in the background to roughly
simulate a large copy in progress. The expectation is that the IO
should have little or no impact on memcachetest which is running
entirely in memory.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Ops memcachetest-0M 22155.00 ( 0.00%) 22180.00 ( 0.11%)
Ops memcachetest-715M 22720.00 ( 0.00%) 22355.00 ( -1.61%)
Ops memcachetest-2385M 3939.00 ( 0.00%) 23450.00 (495.33%)
Ops memcachetest-4055M 3628.00 ( 0.00%) 24341.00 (570.92%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%)
Ops io-duration-2385M 118.00 ( 0.00%) 21.00 ( 82.20%)
Ops io-duration-4055M 162.00 ( 0.00%) 36.00 ( 77.78%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140134.00 ( 0.00%) 18.00 ( 99.99%)
Ops swaptotal-2385M 392438.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 449037.00 ( 0.00%) 27864.00 ( 93.79%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 148031.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 135109.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1529984.00 ( 0.00%) 1530235.00 ( -0.02%)
Ops minorfaults-715M 1794168.00 ( 0.00%) 1613750.00 ( 10.06%)
Ops minorfaults-2385M 1739813.00 ( 0.00%) 1609396.00 ( 7.50%)
Ops minorfaults-4055M 1754460.00 ( 0.00%) 1614810.00 ( 7.96%)
Ops majorfaults-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 185.00 ( 0.00%) 180.00 ( 2.70%)
Ops majorfaults-2385M 24472.00 ( 0.00%) 101.00 ( 99.59%)
Ops majorfaults-4055M 22302.00 ( 0.00%) 229.00 ( 98.97%)
Note how the vanilla kernels performance collapses when there is enough
IO taking place in the background. This drop in performance is part of
what users complain of when they start backups. Note how the swapin and
major fault figures indicate that processes were being pushed to swap
prematurely. With the series applied, there is no noticable performance
drop and while there is still some swap activity, it's tiny.
20 iterations of this test were run in total and averaged. Every 5
iterations, additional IO was generated in the background using dd to
measure how the workload was impacted. The 0M, 715M, 2385M and 4055M
subblock refer to the amount of IO going on in the background at each
iteration. So memcachetest-2385M is reporting how many
transactions/second memcachetest recorded on average over 5 iterations
while there was 2385M of IO going on in the ground. There are six
blocks of information reported here
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
22K/sec to just under 4K/second when there is 2385M of IO going
on in the background. This is one type of performance collapse
users complain about if a large cp or backup starts in the
background
io-duration refers to how long it takes for the background IO to
complete. It's showing that with the patched kernel that the IO
completes faster while not interfering with the memcache
workload
swaptotal is the total amount of swap traffic. With the patched kernel,
the total amount of swapping is much reduced although it is
still not zero.
swapin in this case is an indication as to whether we are swap trashing.
The closer the swapin/swapout ratio is to 1, the worse the
trashing is. Note with the patched kernel that there is no swapin
activity indicating that all the pages swapped were really inactive
unused pages.
minorfaults are just minor faults. An increased number of minor faults
can indicate that page reclaim is unmapping the pages but not
swapping them out before they are faulted back in. With the
patched kernel, there is only a small change in minor faults
majorfaults are just major faults in the target workload and a high
number can indicate that a workload is being prematurely
swapped. With the patched kernel, major faults are much reduced. As
there are no swapin's recorded so it's not being swapped. The likely
explanation is that that libraries or configuration files used by
the workload during startup get paged out by the background IO.
Overall with the series applied, there is no noticable performance drop
due to background IO and while there is still some swap activity, it's
tiny and the lack of swapins imply that the swapped pages were inactive
and unused.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Page Ins 1234608 101892
Page Outs 12446272 11810468
Swap Ins 283406 0
Swap Outs 698469 27882
Direct pages scanned 0 136480
Kswapd pages scanned 6266537 5369364
Kswapd pages reclaimed 1088989 930832
Direct pages reclaimed 0 120901
Kswapd efficiency 17% 17%
Kswapd velocity 5398.371 4635.115
Direct efficiency 100% 88%
Direct velocity 0.000 117.817
Percentage direct scans 0% 2%
Page writes by reclaim 1655843 4009929
Page writes file 957374 3982047
Page writes anon 698469 27882
Page reclaim immediate 5245 1745
Page rescued immediate 0 0
Slabs scanned 33664 25216
Direct inode steals 0 0
Kswapd inode steals 19409 778
Kswapd skipped wait 0 0
THP fault alloc 35 30
THP collapse alloc 472 401
THP splits 27 22
THP fault fallback 0 0
THP collapse fail 0 1
Compaction stalls 0 4
Compaction success 0 0
Compaction failures 0 4
Page migrate success 0 0
Page migrate failure 0 0
Compaction pages isolated 0 0
Compaction migrate scanned 0 0
Compaction free scanned 0 0
Compaction cost 0 0
NUMA PTE updates 0 0
NUMA hint faults 0 0
NUMA hint local faults 0 0
NUMA pages migrated 0 0
AutoNUMA cost 0 0
Unfortunately, note that there is a small amount of direct reclaim due to
kswapd no longer reclaiming the world. ftrace indicates that the direct
reclaim stalls are mostly harmless with the vast bulk of the stalls
incurred by dd
23 tclsh-3367
38 memcachetest-13733
49 memcachetest-12443
57 tee-3368
1541 dd-13826
1981 dd-12539
A consequence of the direct reclaim for dd is that the processes for the
IO workload may show a higher system CPU usage. There is also a risk that
kswapd not reclaiming the world may mean that it stays awake balancing
zones, does not stall on the appropriate events and continually scans
pages it cannot reclaim consuming CPU. This will be visible as continued
high CPU usage but in my own tests I only saw a single spike lasting less
than a second and I did not observe any problems related to reclaim while
running the series on my desktop.
This patch:
The number of pages kswapd can reclaim is bound by the number of pages it
scans which is related to the size of the zone and the scanning priority.
In many cases the priority remains low because it's reset every
SWAP_CLUSTER_MAX reclaimed pages but in the event kswapd scans a large
number of pages it cannot reclaim, it will raise the priority and
potentially discard a large percentage of the zone as sc->nr_to_reclaim is
ULONG_MAX. The user-visible effect is a reclaim "spike" where a large
percentage of memory is suddenly freed. It would be bad enough if this
was just unused memory but because of how anon/file pages are balanced it
is possible that applications get pushed to swap unnecessarily.
This patch limits the number of pages kswapd will reclaim to the high
watermark. Reclaim will still overshoot due to it not being a hard limit
as shrink_lruvec() will ignore the sc.nr_to_reclaim at DEF_PRIORITY but it
prevents kswapd reclaiming the world at higher priorities. The number of
pages it reclaims is not adjusted for high-order allocations as kswapd
will reclaim excessively if it is to balance zones for high-order
allocations.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Tested-by: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:42 +00:00
|
|
|
|
2016-07-28 22:45:43 +00:00
|
|
|
/* Reclaim a number of pages proportional to the number of zones */
|
|
|
|
sc->nr_to_reclaim = 0;
|
2016-07-28 22:46:35 +00:00
|
|
|
for (z = 0; z <= sc->reclaim_idx; z++) {
|
2016-07-28 22:45:43 +00:00
|
|
|
zone = pgdat->node_zones + z;
|
2016-09-01 23:14:55 +00:00
|
|
|
if (!managed_zone(zone))
|
2016-07-28 22:45:43 +00:00
|
|
|
continue;
|
2013-07-03 22:01:54 +00:00
|
|
|
|
2016-07-28 22:45:43 +00:00
|
|
|
sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
|
|
|
|
}
|
2013-07-03 22:01:54 +00:00
|
|
|
|
|
|
|
/*
|
2016-07-28 22:45:43 +00:00
|
|
|
* Historically care was taken to put equal pressure on all zones but
|
|
|
|
* now pressure is applied based on node LRU order.
|
2013-07-03 22:01:54 +00:00
|
|
|
*/
|
2016-07-28 22:46:35 +00:00
|
|
|
shrink_node(pgdat, sc);
|
2013-07-03 22:01:51 +00:00
|
|
|
|
2013-07-03 22:01:54 +00:00
|
|
|
/*
|
2016-07-28 22:45:43 +00:00
|
|
|
* Fragmentation may mean that the system cannot be rebalanced for
|
|
|
|
* high-order allocations. If twice the allocation size has been
|
|
|
|
* reclaimed then recheck watermarks only at order-0 to prevent
|
|
|
|
* excessive reclaim. Assume that a process requested a high-order
|
|
|
|
* can direct reclaim/compact.
|
2013-07-03 22:01:54 +00:00
|
|
|
*/
|
2016-10-07 23:57:53 +00:00
|
|
|
if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
|
2016-07-28 22:45:43 +00:00
|
|
|
sc->order = 0;
|
2013-07-03 22:01:54 +00:00
|
|
|
|
2013-07-03 22:01:45 +00:00
|
|
|
return sc->nr_scanned >= sc->nr_to_reclaim;
|
mm: vmscan: limit the number of pages kswapd reclaims at each priority
This series does not fix all the current known problems with reclaim but
it addresses one important swapping bug when there is background IO.
Changelog since V3
- Drop the slab shrink changes in light of Glaubers series and
discussions highlighted that there were a number of potential
problems with the patch. (mel)
- Rebased to 3.10-rc1
Changelog since V2
- Preserve ratio properly for proportional scanning (kamezawa)
Changelog since V1
- Rename ZONE_DIRTY to ZONE_TAIL_LRU_DIRTY (andi)
- Reformat comment in shrink_page_list (andi)
- Clarify some comments (dhillf)
- Rework how the proportional scanning is preserved
- Add PageReclaim check before kswapd starts writeback
- Reset sc.nr_reclaimed on every full zone scan
Kswapd and page reclaim behaviour has been screwy in one way or the
other for a long time. Very broadly speaking it worked in the far past
because machines were limited in memory so it did not have that many
pages to scan and it stalled congestion_wait() frequently to prevent it
going completely nuts. In recent times it has behaved very
unsatisfactorily with some of the problems compounded by the removal of
stall logic and the introduction of transparent hugepage support with
high-order reclaims.
There are many variations of bugs that are rooted in this area. One
example is reports of a large copy operations or backup causing the
machine to grind to a halt or applications pushed to swap. Sometimes in
low memory situations a large percentage of memory suddenly gets
reclaimed. In other cases an application starts and kswapd hits 100%
CPU usage for prolonged periods of time and so on. There is now talk of
introducing features like an extra free kbytes tunable to work around
aspects of the problem instead of trying to deal with it. It's
compounded by the problem that it can be very workload and machine
specific.
This series aims at addressing some of the worst of these problems
without attempting to fundmentally alter how page reclaim works.
Patches 1-2 limits the number of pages kswapd reclaims while still obeying
the anon/file proportion of the LRUs it should be scanning.
Patches 3-4 control how and when kswapd raises its scanning priority and
deletes the scanning restart logic which is tricky to follow.
Patch 5 notes that it is too easy for kswapd to reach priority 0 when
scanning and then reclaim the world. Down with that sort of thing.
Patch 6 notes that kswapd starts writeback based on scanning priority which
is not necessarily related to dirty pages. It will have kswapd
writeback pages if a number of unqueued dirty pages have been
recently encountered at the tail of the LRU.
Patch 7 notes that sometimes kswapd should stall waiting on IO to complete
to reduce LRU churn and the likelihood that it'll reclaim young
clean pages or push applications to swap. It will cause kswapd
to block on IO if it detects that pages being reclaimed under
writeback are recycling through the LRU before the IO completes.
Patchies 8-9 are cosmetic but balance_pgdat() is easier to follow after they
are applied.
This was tested using memcached+memcachetest while some background IO
was in progress as implemented by the parallel IO tests implement in MM
Tests.
memcachetest benchmarks how many operations/second memcached can service
and it is run multiple times. It starts with no background IO and then
re-runs the test with larger amounts of IO in the background to roughly
simulate a large copy in progress. The expectation is that the IO
should have little or no impact on memcachetest which is running
entirely in memory.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Ops memcachetest-0M 22155.00 ( 0.00%) 22180.00 ( 0.11%)
Ops memcachetest-715M 22720.00 ( 0.00%) 22355.00 ( -1.61%)
Ops memcachetest-2385M 3939.00 ( 0.00%) 23450.00 (495.33%)
Ops memcachetest-4055M 3628.00 ( 0.00%) 24341.00 (570.92%)
Ops io-duration-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops io-duration-715M 12.00 ( 0.00%) 7.00 ( 41.67%)
Ops io-duration-2385M 118.00 ( 0.00%) 21.00 ( 82.20%)
Ops io-duration-4055M 162.00 ( 0.00%) 36.00 ( 77.78%)
Ops swaptotal-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-715M 140134.00 ( 0.00%) 18.00 ( 99.99%)
Ops swaptotal-2385M 392438.00 ( 0.00%) 0.00 ( 0.00%)
Ops swaptotal-4055M 449037.00 ( 0.00%) 27864.00 ( 93.79%)
Ops swapin-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-715M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-2385M 148031.00 ( 0.00%) 0.00 ( 0.00%)
Ops swapin-4055M 135109.00 ( 0.00%) 0.00 ( 0.00%)
Ops minorfaults-0M 1529984.00 ( 0.00%) 1530235.00 ( -0.02%)
Ops minorfaults-715M 1794168.00 ( 0.00%) 1613750.00 ( 10.06%)
Ops minorfaults-2385M 1739813.00 ( 0.00%) 1609396.00 ( 7.50%)
Ops minorfaults-4055M 1754460.00 ( 0.00%) 1614810.00 ( 7.96%)
Ops majorfaults-0M 0.00 ( 0.00%) 0.00 ( 0.00%)
Ops majorfaults-715M 185.00 ( 0.00%) 180.00 ( 2.70%)
Ops majorfaults-2385M 24472.00 ( 0.00%) 101.00 ( 99.59%)
Ops majorfaults-4055M 22302.00 ( 0.00%) 229.00 ( 98.97%)
Note how the vanilla kernels performance collapses when there is enough
IO taking place in the background. This drop in performance is part of
what users complain of when they start backups. Note how the swapin and
major fault figures indicate that processes were being pushed to swap
prematurely. With the series applied, there is no noticable performance
drop and while there is still some swap activity, it's tiny.
20 iterations of this test were run in total and averaged. Every 5
iterations, additional IO was generated in the background using dd to
measure how the workload was impacted. The 0M, 715M, 2385M and 4055M
subblock refer to the amount of IO going on in the background at each
iteration. So memcachetest-2385M is reporting how many
transactions/second memcachetest recorded on average over 5 iterations
while there was 2385M of IO going on in the ground. There are six
blocks of information reported here
memcachetest is the transactions/second reported by memcachetest. In
the vanilla kernel note that performance drops from around
22K/sec to just under 4K/second when there is 2385M of IO going
on in the background. This is one type of performance collapse
users complain about if a large cp or backup starts in the
background
io-duration refers to how long it takes for the background IO to
complete. It's showing that with the patched kernel that the IO
completes faster while not interfering with the memcache
workload
swaptotal is the total amount of swap traffic. With the patched kernel,
the total amount of swapping is much reduced although it is
still not zero.
swapin in this case is an indication as to whether we are swap trashing.
The closer the swapin/swapout ratio is to 1, the worse the
trashing is. Note with the patched kernel that there is no swapin
activity indicating that all the pages swapped were really inactive
unused pages.
minorfaults are just minor faults. An increased number of minor faults
can indicate that page reclaim is unmapping the pages but not
swapping them out before they are faulted back in. With the
patched kernel, there is only a small change in minor faults
majorfaults are just major faults in the target workload and a high
number can indicate that a workload is being prematurely
swapped. With the patched kernel, major faults are much reduced. As
there are no swapin's recorded so it's not being swapped. The likely
explanation is that that libraries or configuration files used by
the workload during startup get paged out by the background IO.
Overall with the series applied, there is no noticable performance drop
due to background IO and while there is still some swap activity, it's
tiny and the lack of swapins imply that the swapped pages were inactive
and unused.
3.10.0-rc1 3.10.0-rc1
vanilla lessdisrupt-v4
Page Ins 1234608 101892
Page Outs 12446272 11810468
Swap Ins 283406 0
Swap Outs 698469 27882
Direct pages scanned 0 136480
Kswapd pages scanned 6266537 5369364
Kswapd pages reclaimed 1088989 930832
Direct pages reclaimed 0 120901
Kswapd efficiency 17% 17%
Kswapd velocity 5398.371 4635.115
Direct efficiency 100% 88%
Direct velocity 0.000 117.817
Percentage direct scans 0% 2%
Page writes by reclaim 1655843 4009929
Page writes file 957374 3982047
Page writes anon 698469 27882
Page reclaim immediate 5245 1745
Page rescued immediate 0 0
Slabs scanned 33664 25216
Direct inode steals 0 0
Kswapd inode steals 19409 778
Kswapd skipped wait 0 0
THP fault alloc 35 30
THP collapse alloc 472 401
THP splits 27 22
THP fault fallback 0 0
THP collapse fail 0 1
Compaction stalls 0 4
Compaction success 0 0
Compaction failures 0 4
Page migrate success 0 0
Page migrate failure 0 0
Compaction pages isolated 0 0
Compaction migrate scanned 0 0
Compaction free scanned 0 0
Compaction cost 0 0
NUMA PTE updates 0 0
NUMA hint faults 0 0
NUMA hint local faults 0 0
NUMA pages migrated 0 0
AutoNUMA cost 0 0
Unfortunately, note that there is a small amount of direct reclaim due to
kswapd no longer reclaiming the world. ftrace indicates that the direct
reclaim stalls are mostly harmless with the vast bulk of the stalls
incurred by dd
23 tclsh-3367
38 memcachetest-13733
49 memcachetest-12443
57 tee-3368
1541 dd-13826
1981 dd-12539
A consequence of the direct reclaim for dd is that the processes for the
IO workload may show a higher system CPU usage. There is also a risk that
kswapd not reclaiming the world may mean that it stays awake balancing
zones, does not stall on the appropriate events and continually scans
pages it cannot reclaim consuming CPU. This will be visible as continued
high CPU usage but in my own tests I only saw a single spike lasting less
than a second and I did not observe any problems related to reclaim while
running the series on my desktop.
This patch:
The number of pages kswapd can reclaim is bound by the number of pages it
scans which is related to the size of the zone and the scanning priority.
In many cases the priority remains low because it's reset every
SWAP_CLUSTER_MAX reclaimed pages but in the event kswapd scans a large
number of pages it cannot reclaim, it will raise the priority and
potentially discard a large percentage of the zone as sc->nr_to_reclaim is
ULONG_MAX. The user-visible effect is a reclaim "spike" where a large
percentage of memory is suddenly freed. It would be bad enough if this
was just unused memory but because of how anon/file pages are balanced it
is possible that applications get pushed to swap unnecessarily.
This patch limits the number of pages kswapd will reclaim to the high
watermark. Reclaim will still overshoot due to it not being a hard limit
as shrink_lruvec() will ignore the sc.nr_to_reclaim at DEF_PRIORITY but it
prevents kswapd reclaiming the world at higher priorities. The number of
pages it reclaims is not adjusted for high-order allocations as kswapd
will reclaim excessively if it is to balance zones for high-order
allocations.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Tested-by: Zlatko Calusic <zcalusic@bitsync.net>
Cc: dormando <dormando@rydia.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:01:42 +00:00
|
|
|
}
|
|
|
|
|
2021-06-29 02:42:21 +00:00
|
|
|
/* Page allocator PCP high watermark is lowered if reclaim is active. */
|
|
|
|
static inline void
|
|
|
|
update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
struct zone *zone;
|
|
|
|
|
|
|
|
for (i = 0; i <= highest_zoneidx; i++) {
|
|
|
|
zone = pgdat->node_zones + i;
|
|
|
|
|
|
|
|
if (!managed_zone(zone))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (active)
|
|
|
|
set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
|
|
|
|
else
|
|
|
|
clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void
|
|
|
|
set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
|
|
|
|
{
|
|
|
|
update_reclaim_active(pgdat, highest_zoneidx, true);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void
|
|
|
|
clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
|
|
|
|
{
|
|
|
|
update_reclaim_active(pgdat, highest_zoneidx, false);
|
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2016-07-28 22:45:43 +00:00
|
|
|
* For kswapd, balance_pgdat() will reclaim pages across a node from zones
|
|
|
|
* that are eligible for use by the caller until at least one zone is
|
|
|
|
* balanced.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2016-07-28 22:45:43 +00:00
|
|
|
* Returns the order kswapd finished reclaiming at.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* kswapd scans the zones in the highmem->normal->dma direction. It skips
|
2009-06-16 22:32:12 +00:00
|
|
|
* zones which have free_pages > high_wmark_pages(zone), but once a zone is
|
2019-03-05 23:46:22 +00:00
|
|
|
* found to have free_pages <= high_wmark_pages(zone), any page in that zone
|
2016-07-28 22:45:43 +00:00
|
|
|
* or lower is eligible for reclaim until at least one usable zone is
|
|
|
|
* balanced.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
|
|
|
int i;
|
2013-09-24 22:27:41 +00:00
|
|
|
unsigned long nr_soft_reclaimed;
|
|
|
|
unsigned long nr_soft_scanned;
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:06:27 +00:00
|
|
|
unsigned long pflags;
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
unsigned long nr_boost_reclaim;
|
|
|
|
unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
|
|
|
|
bool boosted;
|
2016-07-28 22:45:43 +00:00
|
|
|
struct zone *zone;
|
2006-03-22 08:08:18 +00:00
|
|
|
struct scan_control sc = {
|
|
|
|
.gfp_mask = GFP_KERNEL,
|
2014-08-06 23:06:19 +00:00
|
|
|
.order = order,
|
2009-03-31 22:19:30 +00:00
|
|
|
.may_unmap = 1,
|
2006-03-22 08:08:18 +00:00
|
|
|
};
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(current, &sc.reclaim_state);
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:06:27 +00:00
|
|
|
psi_memstall_enter(&pflags);
|
2021-09-02 21:52:58 +00:00
|
|
|
__fs_reclaim_acquire(_THIS_IP_);
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
|
2006-06-30 08:55:45 +00:00
|
|
|
count_vm_event(PAGEOUTRUN);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
/*
|
|
|
|
* Account for the reclaim boost. Note that the zone boost is left in
|
|
|
|
* place so that parallel allocations that are near the watermark will
|
|
|
|
* stall or direct reclaim until kswapd is finished.
|
|
|
|
*/
|
|
|
|
nr_boost_reclaim = 0;
|
2020-06-03 22:59:01 +00:00
|
|
|
for (i = 0; i <= highest_zoneidx; i++) {
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
zone = pgdat->node_zones + i;
|
|
|
|
if (!managed_zone(zone))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
nr_boost_reclaim += zone->watermark_boost;
|
|
|
|
zone_boosts[i] = zone->watermark_boost;
|
|
|
|
}
|
|
|
|
boosted = nr_boost_reclaim;
|
|
|
|
|
|
|
|
restart:
|
2021-06-29 02:42:21 +00:00
|
|
|
set_reclaim_active(pgdat, highest_zoneidx);
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
sc.priority = DEF_PRIORITY;
|
2012-05-29 22:06:57 +00:00
|
|
|
do {
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
unsigned long nr_reclaimed = sc.nr_reclaimed;
|
2013-07-03 22:01:45 +00:00
|
|
|
bool raise_priority = true;
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
bool balanced;
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
bool ret;
|
2013-07-03 22:01:45 +00:00
|
|
|
|
2020-06-03 22:59:01 +00:00
|
|
|
sc.reclaim_idx = highest_zoneidx;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2016-07-28 22:45:59 +00:00
|
|
|
/*
|
2016-07-28 22:46:44 +00:00
|
|
|
* If the number of buffer_heads exceeds the maximum allowed
|
|
|
|
* then consider reclaiming from all zones. This has a dual
|
|
|
|
* purpose -- on 64-bit systems it is expected that
|
|
|
|
* buffer_heads are stripped during active rotation. On 32-bit
|
|
|
|
* systems, highmem pages can pin lowmem memory and shrinking
|
|
|
|
* buffers can relieve lowmem pressure. Reclaim may still not
|
|
|
|
* go ahead if all eligible zones for the original allocation
|
|
|
|
* request are balanced to avoid excessive reclaim from kswapd.
|
2016-07-28 22:45:59 +00:00
|
|
|
*/
|
|
|
|
if (buffer_heads_over_limit) {
|
|
|
|
for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
|
|
|
|
zone = pgdat->node_zones + i;
|
2016-09-01 23:14:55 +00:00
|
|
|
if (!managed_zone(zone))
|
2016-07-28 22:45:59 +00:00
|
|
|
continue;
|
2012-03-21 23:34:00 +00:00
|
|
|
|
2016-07-28 22:46:35 +00:00
|
|
|
sc.reclaim_idx = i;
|
2006-12-07 04:32:01 +00:00
|
|
|
break;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
}
|
2013-02-23 00:32:34 +00:00
|
|
|
|
2016-07-28 22:45:59 +00:00
|
|
|
/*
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
* If the pgdat is imbalanced then ignore boosting and preserve
|
|
|
|
* the watermarks for a later time and restart. Note that the
|
|
|
|
* zone watermarks will be still reset at the end of balancing
|
|
|
|
* on the grounds that the normal reclaim should be enough to
|
|
|
|
* re-evaluate if boosting is required when kswapd next wakes.
|
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
if (!balanced && nr_boost_reclaim) {
|
|
|
|
nr_boost_reclaim = 0;
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If boosting is not active then only reclaim if there are no
|
|
|
|
* eligible zones. Note that sc.reclaim_idx is not used as
|
|
|
|
* buffer_heads_over_limit may have adjusted it.
|
2016-07-28 22:45:59 +00:00
|
|
|
*/
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
if (!nr_boost_reclaim && balanced)
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
goto out;
|
2006-12-07 04:32:01 +00:00
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
/* Limit the priority of boosting to avoid reclaim writeback */
|
|
|
|
if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
|
|
|
|
raise_priority = false;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Do not writeback or swap pages for boosted reclaim. The
|
|
|
|
* intent is to relieve pressure not issue sub-optimal IO
|
|
|
|
* from reclaim context. If no pages are reclaimed, the
|
|
|
|
* reclaim will be aborted.
|
|
|
|
*/
|
|
|
|
sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
|
|
|
|
sc.may_swap = !nr_boost_reclaim;
|
|
|
|
|
2016-07-28 22:45:43 +00:00
|
|
|
/*
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
* Do some background aging, to give pages a chance to be
|
|
|
|
* referenced before reclaiming. All pages are rotated
|
|
|
|
* regardless of classzone as this is about consistent aging.
|
2016-07-28 22:45:43 +00:00
|
|
|
*/
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 08:00:03 +00:00
|
|
|
kswapd_age_node(pgdat, &sc);
|
2016-07-28 22:45:43 +00:00
|
|
|
|
2013-07-03 22:01:53 +00:00
|
|
|
/*
|
|
|
|
* If we're getting trouble reclaiming, start doing writepage
|
|
|
|
* even in laptop mode.
|
|
|
|
*/
|
2017-05-03 21:51:57 +00:00
|
|
|
if (sc.priority < DEF_PRIORITY - 2)
|
2013-07-03 22:01:53 +00:00
|
|
|
sc.may_writepage = 1;
|
|
|
|
|
2016-07-28 22:45:43 +00:00
|
|
|
/* Call soft limit reclaim before calling shrink_node. */
|
|
|
|
sc.nr_scanned = 0;
|
|
|
|
nr_soft_scanned = 0;
|
2016-07-28 22:46:05 +00:00
|
|
|
nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
|
2016-07-28 22:45:43 +00:00
|
|
|
sc.gfp_mask, &nr_soft_scanned);
|
|
|
|
sc.nr_reclaimed += nr_soft_reclaimed;
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2016-07-28 22:45:43 +00:00
|
|
|
* There should be no need to raise the scanning priority if
|
|
|
|
* enough pages are already being scanned that that high
|
|
|
|
* watermark would be met at 100% efficiency.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2016-07-28 22:46:35 +00:00
|
|
|
if (kswapd_shrink_node(pgdat, &sc))
|
2016-07-28 22:45:43 +00:00
|
|
|
raise_priority = false;
|
2012-07-31 23:44:35 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* If the low watermark is met there is no need for processes
|
|
|
|
* to be throttled on pfmemalloc_wait as they should not be
|
|
|
|
* able to safely make forward progress. Wake them
|
|
|
|
*/
|
|
|
|
if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
allow_direct_reclaim(pgdat))
|
2015-02-11 23:25:12 +00:00
|
|
|
wake_up_all(&pgdat->pfmemalloc_wait);
|
2012-07-31 23:44:35 +00:00
|
|
|
|
2013-07-03 22:01:45 +00:00
|
|
|
/* Check if kswapd should be suspending */
|
2021-09-02 21:52:58 +00:00
|
|
|
__fs_reclaim_release(_THIS_IP_);
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
ret = try_to_freeze();
|
2021-09-02 21:52:58 +00:00
|
|
|
__fs_reclaim_acquire(_THIS_IP_);
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
if (ret || kthread_should_stop())
|
2013-07-03 22:01:45 +00:00
|
|
|
break;
|
[PATCH] swsusp: Improve handling of highmem
Currently swsusp saves the contents of highmem pages by copying them to the
normal zone which is quite inefficient (eg. it requires two normal pages
to be used for saving one highmem page). This may be improved by using
highmem for saving the contents of saveable highmem pages.
Namely, during the suspend phase of the suspend-resume cycle we try to
allocate as many free highmem pages as there are saveable highmem pages.
If there are not enough highmem image pages to store the contents of all of
the saveable highmem pages, some of them will be stored in the "normal"
memory. Next, we allocate as many free "normal" pages as needed to store
the (remaining) image data. We use a memory bitmap to mark the allocated
free pages (ie. highmem as well as "normal" image pages).
Now, we use another memory bitmap to mark all of the saveable pages
(highmem as well as "normal") and the contents of the saveable pages are
copied into the image pages. Then, the second bitmap is used to save the
pfns corresponding to the saveable pages and the first one is used to save
their data.
During the resume phase the pfns of the pages that were saveable during the
suspend are loaded from the image and used to mark the "unsafe" page
frames. Next, we try to allocate as many free highmem page frames as to
load all of the image data that had been in the highmem before the suspend
and we allocate so many free "normal" page frames that the total number of
allocated free pages (highmem and "normal") is equal to the size of the
image. While doing this we have to make sure that there will be some extra
free "normal" and "safe" page frames for two lists of PBEs constructed
later.
Now, the image data are loaded, if possible, into their "original" page
frames. The image data that cannot be written into their "original" page
frames are loaded into "safe" page frames and their "original" kernel
virtual addresses, as well as the addresses of the "safe" pages containing
their copies, are stored in one of two lists of PBEs.
One list of PBEs is for the copies of "normal" suspend pages (ie. "normal"
pages that were saveable during the suspend) and it is used in the same way
as previously (ie. by the architecture-dependent parts of swsusp). The
other list of PBEs is for the copies of highmem suspend pages. The pages
in this list are restored (in a reversible way) right before the
arch-dependent code is called.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Pavel Machek <pavel@ucw.cz>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 04:34:18 +00:00
|
|
|
|
2009-01-06 22:40:33 +00:00
|
|
|
/*
|
2013-07-03 22:01:45 +00:00
|
|
|
* Raise priority if scanning rate is too low or there was no
|
|
|
|
* progress in reclaiming pages
|
2009-01-06 22:40:33 +00:00
|
|
|
*/
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If reclaim made no progress for a boost, stop reclaim as
|
|
|
|
* IO cannot be queued and it could be an infinite loop in
|
|
|
|
* extreme circumstances.
|
|
|
|
*/
|
|
|
|
if (nr_boost_reclaim && !nr_reclaimed)
|
|
|
|
break;
|
|
|
|
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
if (raise_priority || !nr_reclaimed)
|
2013-07-03 22:01:45 +00:00
|
|
|
sc.priority--;
|
2016-07-28 22:45:43 +00:00
|
|
|
} while (sc.priority >= 1);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
mm: fix 100% CPU kswapd busyloop on unreclaimable nodes
Patch series "mm: kswapd spinning on unreclaimable nodes - fixes and
cleanups".
Jia reported a scenario in which the kswapd of a node indefinitely spins
at 100% CPU usage. We have seen similar cases at Facebook.
The kernel's current method of judging its ability to reclaim a node (or
whether to back off and sleep) is based on the amount of scanned pages
in proportion to the amount of reclaimable pages. In Jia's and our
scenarios, there are no reclaimable pages in the node, however, and the
condition for backing off is never met. Kswapd busyloops in an attempt
to restore the watermarks while having nothing to work with.
This series reworks the definition of an unreclaimable node based not on
scanning but on whether kswapd is able to actually reclaim pages in
MAX_RECLAIM_RETRIES (16) consecutive runs. This is the same criteria
the page allocator uses for giving up on direct reclaim and invoking the
OOM killer. If it cannot free any pages, kswapd will go to sleep and
leave further attempts to direct reclaim invocations, which will either
make progress and re-enable kswapd, or invoke the OOM killer.
Patch #1 fixes the immediate problem Jia reported, the remainder are
smaller fixlets, cleanups, and overall phasing out of the old method.
Patch #6 is the odd one out. It's a nice cleanup to get_scan_count(),
and directly related to #5, but in itself not relevant to the series.
If the whole series is too ambitious for 4.11, I would consider the
first three patches fixes, the rest cleanups.
This patch (of 9):
Jia He reports a problem with kswapd spinning at 100% CPU when
requesting more hugepages than memory available in the system:
$ echo 4000 >/proc/sys/vm/nr_hugepages
top - 13:42:59 up 3:37, 1 user, load average: 1.09, 1.03, 1.01
Tasks: 1 total, 1 running, 0 sleeping, 0 stopped, 0 zombie
%Cpu(s): 0.0 us, 12.5 sy, 0.0 ni, 85.5 id, 2.0 wa, 0.0 hi, 0.0 si, 0.0 st
KiB Mem: 31371520 total, 30915136 used, 456384 free, 320 buffers
KiB Swap: 6284224 total, 115712 used, 6168512 free. 48192 cached Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
76 root 20 0 0 0 0 R 100.0 0.000 217:17.29 kswapd3
At that time, there are no reclaimable pages left in the node, but as
kswapd fails to restore the high watermarks it refuses to go to sleep.
Kswapd needs to back away from nodes that fail to balance. Up until
commit 1d82de618ddd ("mm, vmscan: make kswapd reclaim in terms of
nodes") kswapd had such a mechanism. It considered zones whose
theoretically reclaimable pages it had reclaimed six times over as
unreclaimable and backed away from them. This guard was erroneously
removed as the patch changed the definition of a balanced node.
However, simply restoring this code wouldn't help in the case reported
here: there *are* no reclaimable pages that could be scanned until the
threshold is met. Kswapd would stay awake anyway.
Introduce a new and much simpler way of backing off. If kswapd runs
through MAX_RECLAIM_RETRIES (16) cycles without reclaiming a single
page, make it back off from the node. This is the same number of shots
direct reclaim takes before declaring OOM. Kswapd will go to sleep on
that node until a direct reclaimer manages to reclaim some pages, thus
proving the node reclaimable again.
[hannes@cmpxchg.org: check kswapd failure against the cumulative nr_reclaimed count]
Link: http://lkml.kernel.org/r/20170306162410.GB2090@cmpxchg.org
[shakeelb@google.com: fix condition for throttle_direct_reclaim]
Link: http://lkml.kernel.org/r/20170314183228.20152-1-shakeelb@google.com
Link: http://lkml.kernel.org/r/20170228214007.5621-2-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Reported-by: Jia He <hejianet@gmail.com>
Tested-by: Jia He <hejianet@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:51:51 +00:00
|
|
|
if (!sc.nr_reclaimed)
|
|
|
|
pgdat->kswapd_failures++;
|
|
|
|
|
2013-07-03 22:01:45 +00:00
|
|
|
out:
|
2021-06-29 02:42:21 +00:00
|
|
|
clear_reclaim_active(pgdat, highest_zoneidx);
|
|
|
|
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
/* If reclaim was boosted, account for the reclaim done in this pass */
|
|
|
|
if (boosted) {
|
|
|
|
unsigned long flags;
|
|
|
|
|
2020-06-03 22:59:01 +00:00
|
|
|
for (i = 0; i <= highest_zoneidx; i++) {
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
if (!zone_boosts[i])
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/* Increments are under the zone lock */
|
|
|
|
zone = pgdat->node_zones + i;
|
|
|
|
spin_lock_irqsave(&zone->lock, flags);
|
|
|
|
zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
|
|
|
|
spin_unlock_irqrestore(&zone->lock, flags);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* As there is now likely space, wakeup kcompact to defragment
|
|
|
|
* pageblocks.
|
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
|
mm: reclaim small amounts of memory when an external fragmentation event occurs
An external fragmentation event was previously described as
When the page allocator fragments memory, it records the event using
the mm_page_alloc_extfrag event. If the fallback_order is smaller
than a pageblock order (order-9 on 64-bit x86) then it's considered
an event that will cause external fragmentation issues in the future.
The kernel reduces the probability of such events by increasing the
watermark sizes by calling set_recommended_min_free_kbytes early in the
lifetime of the system. This works reasonably well in general but if
there are enough sparsely populated pageblocks then the problem can still
occur as enough memory is free overall and kswapd stays asleep.
This patch introduces a watermark_boost_factor sysctl that allows a zone
watermark to be temporarily boosted when an external fragmentation causing
events occurs. The boosting will stall allocations that would decrease
free memory below the boosted low watermark and kswapd is woken if the
calling context allows to reclaim an amount of memory relative to the size
of the high watermark and the watermark_boost_factor until the boost is
cleared. When kswapd finishes, it wakes kcompactd at the pageblock order
to clean some of the pageblocks that may have been affected by the
fragmentation event. kswapd avoids any writeback, slab shrinkage and swap
from reclaim context during this operation to avoid excessive system
disruption in the name of fragmentation avoidance. Care is taken so that
kswapd will do normal reclaim work if the system is really low on memory.
This was evaluated using the same workloads as "mm, page_alloc: Spread
allocations across zones before introducing fragmentation".
1-socket Skylake machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 1 THP allocating thread
--------------------------------------
4.20-rc3 extfrag events < order 9: 804694
4.20-rc3+patch: 408912 (49% reduction)
4.20-rc3+patch1-4: 18421 (98% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-1 653.58 ( 0.00%) 652.71 ( 0.13%)
Amean fault-huge-1 0.00 ( 0.00%) 178.93 * -99.00%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 0.00 ( 0.00%) 5.12 ( 100.00%)
Note that external fragmentation causing events are massively reduced by
this path whether in comparison to the previous kernel or the vanilla
kernel. The fault latency for huge pages appears to be increased but that
is only because THP allocations were successful with the patch applied.
1-socket Skylake machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 291392
4.20-rc3+patch: 191187 (34% reduction)
4.20-rc3+patch1-4: 13464 (95% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Min fault-base-1 912.00 ( 0.00%) 905.00 ( 0.77%)
Min fault-huge-1 127.00 ( 0.00%) 135.00 ( -6.30%)
Amean fault-base-1 1467.55 ( 0.00%) 1481.67 ( -0.96%)
Amean fault-huge-1 1127.11 ( 0.00%) 1063.88 * 5.61%*
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-1 77.64 ( 0.00%) 83.46 ( 7.49%)
As before, massive reduction in external fragmentation events, some jitter
on latencies and an increase in THP allocation success rates.
2-socket Haswell machine
config-global-dhp__workload_thpfioscale XFS (no special madvise)
4 fio threads, 5 THP allocating threads
----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 215698
4.20-rc3+patch: 200210 (7% reduction)
4.20-rc3+patch1-4: 14263 (93% reduction)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 1346.45 ( 0.00%) 1306.87 ( 2.94%)
Amean fault-huge-5 3418.60 ( 0.00%) 1348.94 ( 60.54%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 0.78 ( 0.00%) 7.91 ( 910.64%)
There is a 93% reduction in fragmentation causing events, there is a big
reduction in the huge page fault latency and allocation success rate is
higher.
2-socket Haswell machine
global-dhp__workload_thpfioscale-madvhugepage-xfs (MADV_HUGEPAGE)
-----------------------------------------------------------------
4.20-rc3 extfrag events < order 9: 166352
4.20-rc3+patch: 147463 (11% reduction)
4.20-rc3+patch1-4: 11095 (93% reduction)
thpfioscale Fault Latencies
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Amean fault-base-5 6217.43 ( 0.00%) 7419.67 * -19.34%*
Amean fault-huge-5 3163.33 ( 0.00%) 3263.80 ( -3.18%)
4.20.0-rc3 4.20.0-rc3
lowzone-v5r8 boost-v5r8
Percentage huge-5 95.14 ( 0.00%) 87.98 ( -7.53%)
There is a large reduction in fragmentation events with some jitter around
the latencies and success rates. As before, the high THP allocation
success rate does mean the system is under a lot of pressure. However, as
the fragmentation events are reduced, it would be expected that the
long-term allocation success rate would be higher.
Link: http://lkml.kernel.org/r/20181123114528.28802-5-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 08:35:52 +00:00
|
|
|
}
|
|
|
|
|
2017-05-03 21:55:03 +00:00
|
|
|
snapshot_refaults(NULL, pgdat);
|
2021-09-02 21:52:58 +00:00
|
|
|
__fs_reclaim_release(_THIS_IP_);
|
psi: pressure stall information for CPU, memory, and IO
When systems are overcommitted and resources become contended, it's hard
to tell exactly the impact this has on workload productivity, or how close
the system is to lockups and OOM kills. In particular, when machines work
multiple jobs concurrently, the impact of overcommit in terms of latency
and throughput on the individual job can be enormous.
In order to maximize hardware utilization without sacrificing individual
job health or risk complete machine lockups, this patch implements a way
to quantify resource pressure in the system.
A kernel built with CONFIG_PSI=y creates files in /proc/pressure/ that
expose the percentage of time the system is stalled on CPU, memory, or IO,
respectively. Stall states are aggregate versions of the per-task delay
accounting delays:
cpu: some tasks are runnable but not executing on a CPU
memory: tasks are reclaiming, or waiting for swapin or thrashing cache
io: tasks are waiting for io completions
These percentages of walltime can be thought of as pressure percentages,
and they give a general sense of system health and productivity loss
incurred by resource overcommit. They can also indicate when the system
is approaching lockup scenarios and OOMs.
To do this, psi keeps track of the task states associated with each CPU
and samples the time they spend in stall states. Every 2 seconds, the
samples are averaged across CPUs - weighted by the CPUs' non-idle time to
eliminate artifacts from unused CPUs - and translated into percentages of
walltime. A running average of those percentages is maintained over 10s,
1m, and 5m periods (similar to the loadaverage).
[hannes@cmpxchg.org: doc fixlet, per Randy]
Link: http://lkml.kernel.org/r/20180828205625.GA14030@cmpxchg.org
[hannes@cmpxchg.org: code optimization]
Link: http://lkml.kernel.org/r/20180907175015.GA8479@cmpxchg.org
[hannes@cmpxchg.org: rename psi_clock() to psi_update_work(), per Peter]
Link: http://lkml.kernel.org/r/20180907145404.GB11088@cmpxchg.org
[hannes@cmpxchg.org: fix build]
Link: http://lkml.kernel.org/r/20180913014222.GA2370@cmpxchg.org
Link: http://lkml.kernel.org/r/20180828172258.3185-9-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:06:27 +00:00
|
|
|
psi_memstall_leave(&pflags);
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(current, NULL);
|
2019-07-16 23:26:09 +00:00
|
|
|
|
2011-01-13 23:46:22 +00:00
|
|
|
/*
|
2016-07-28 22:45:43 +00:00
|
|
|
* Return the order kswapd stopped reclaiming at as
|
|
|
|
* prepare_kswapd_sleep() takes it into account. If another caller
|
|
|
|
* entered the allocator slow path while kswapd was awake, order will
|
|
|
|
* remain at the higher level.
|
2011-01-13 23:46:22 +00:00
|
|
|
*/
|
2016-07-28 22:45:43 +00:00
|
|
|
return sc.order;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
/*
|
2020-06-03 22:59:01 +00:00
|
|
|
* The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
|
|
|
|
* be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
|
|
|
|
* not a valid index then either kswapd runs for first time or kswapd couldn't
|
|
|
|
* sleep after previous reclaim attempt (node is still unbalanced). In that
|
|
|
|
* case return the zone index of the previous kswapd reclaim cycle.
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
|
|
|
|
enum zone_type prev_highest_zoneidx)
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
{
|
2020-06-03 22:59:01 +00:00
|
|
|
enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
|
2020-04-02 04:10:12 +00:00
|
|
|
|
2020-06-03 22:59:01 +00:00
|
|
|
return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
}
|
|
|
|
|
2016-07-28 22:45:49 +00:00
|
|
|
static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
|
2020-06-03 22:59:01 +00:00
|
|
|
unsigned int highest_zoneidx)
|
2011-01-13 23:45:50 +00:00
|
|
|
{
|
|
|
|
long remaining = 0;
|
|
|
|
DEFINE_WAIT(wait);
|
|
|
|
|
|
|
|
if (freezing(current) || kthread_should_stop())
|
|
|
|
return;
|
|
|
|
|
|
|
|
prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
|
|
|
|
|
mm, vmscan: fix zone balance check in prepare_kswapd_sleep
Patch series "Reduce amount of time kswapd sleeps prematurely", v2.
The series is unusual in that the first patch fixes one problem and
introduces other issues that are noted in the changelog. Patch 2 makes
a minor modification that is worth considering on its own but leaves the
kernel in a state where it behaves badly. It's not until patch 3 that
there is an improvement against baseline.
This was mostly motivated by examining Chris Mason's "simoop" benchmark
which puts the VM under similar pressure to HADOOP. It has been
reported that the benchmark has regressed severely during the last
number of releases. While I cannot reproduce all the same problems
Chris experienced due to hardware limitations, there was a number of
problems on a 2-socket machine with a single disk.
simoop latencies
4.11.0-rc1 4.11.0-rc1
vanilla keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 125765.57 ( 49.08%)
These are latencies. Read/write are threads reading fixed-size random
blocks from a simulated database. The allocation latency is mmaping and
faulting regions of memory. The p50, 95 and p99 reports the worst
latencies for 50% of the samples, 95% and 99% respectively.
For example, the report indicates that while the test was running 99% of
writes completed 99.75% faster. It's worth noting that on a UMA machine
that no difference in performance with simoop was observed so milage
will vary.
It's noted that there is a slight impact to read latencies but it's
mostly due to IO scheduler decisions and offset by the large reduction
in other latencies.
This patch (of 3):
The check in prepare_kswapd_sleep needs to match the one in
balance_pgdat since the latter will return as soon as any one of the
zones in the classzone is above the watermark. This is specially
important for higher order allocations since balance_pgdat will
typically reset the order to zero relying on compaction to create the
higher order pages. Without this patch, prepare_kswapd_sleep fails to
wake up kcompactd since the zone balance check fails.
It was first reported against 4.9.7 that kswapd is failing to wake up
kcompactd due to a mismatch in the zone balance check between
balance_pgdat() and prepare_kswapd_sleep().
balance_pgdat() returns as soon as a single zone satisfies the
allocation but prepare_kswapd_sleep() requires all zones to do +the
same. This causes prepare_kswapd_sleep() to never succeed except in the
order == 0 case and consequently, wakeup_kcompactd() is never called.
For the machine that originally motivated this patch, the state of
compaction from /proc/vmstat looked this way after a day and a half +of
uptime:
compact_migrate_scanned 240496
compact_free_scanned 76238632
compact_isolated 123472
compact_stall 1791
compact_fail 29
compact_success 1762
compact_daemon_wake 0
After applying the patch and about 10 hours of uptime the state looks
like this:
compact_migrate_scanned 59927299
compact_free_scanned 2021075136
compact_isolated 640926
compact_stall 4
compact_fail 2
compact_success 2
compact_daemon_wake 5160
Further notes from Mel that motivated him to pick this patch up and
resend it;
It was observed for the simoop workload (pressures the VM similar to
HADOOP) that kswapd was failing to keep ahead of direct reclaim. The
investigation noted that there was a need to rationalise kswapd
decisions to reclaim with kswapd decisions to sleep. With this patch on
a 2-socket box, there was a 49% reduction in direct reclaim scanning.
However, the impact otherwise is extremely negative. Kswapd reclaim
efficiency dropped from 98% to 76%. simoop has three latency-related
metrics for read, write and allocation (an anonymous mmap and fault).
4.11.0-rc1 4.11.0-rc1
vanilla fixcheck-v2
Amean p50-Read 21670074.18 ( 0.00%) 20464344.18 ( 5.56%)
Amean p95-Read 25456267.64 ( 0.00%) 25721423.64 ( -1.04%)
Amean p99-Read 29369064.73 ( 0.00%) 30174230.76 ( -2.74%)
Amean p50-Write 1390.30 ( 0.00%) 1395.28 ( -0.36%)
Amean p95-Write 412901.57 ( 0.00%) 37737.74 ( 90.86%)
Amean p99-Write 6668722.09 ( 0.00%) 666489.04 ( 90.01%)
Amean p50-Allocation 78714.31 ( 0.00%) 86286.22 ( -9.62%)
Amean p95-Allocation 175533.51 ( 0.00%) 351812.27 (-100.42%)
Amean p99-Allocation 247003.02 ( 0.00%) 6291171.56 (-2447.00%)
Of greater concern is that the patch causes swapping and page writes
from kswapd context rose from 0 pages to 4189753 pages during the hour
the workload ran for. By and large, the patch has very bad behaviour
but easily missed as the impact on a UMA machine is negligible.
This patch is included with the data in case a bisection leads to this
area. This patch is also a pre-requisite for the rest of the series.
Link: http://lkml.kernel.org/r/20170309075657.25121-2-mgorman@techsingularity.net
Signed-off-by: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:38 +00:00
|
|
|
/*
|
|
|
|
* Try to sleep for a short interval. Note that kcompactd will only be
|
|
|
|
* woken if it is possible to sleep for a short interval. This is
|
|
|
|
* deliberate on the assumption that if reclaim cannot keep an
|
|
|
|
* eligible zone balanced that it's also unlikely that compaction will
|
|
|
|
* succeed.
|
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
|
mm: wake kcompactd before kswapd's short sleep
When kswapd goes to sleep it checks if the node is balanced and at first
it sleeps only for HZ/10 time, then rechecks if the node is still
balanced and nobody has woken it during the initial sleep. Only then it
goes fully sleep until an allocation slowpath wakes it up again.
For higher-order allocations, waking up kcompactd is done only before
the full sleep. This turns out to be an issue in case another
high-order allocation fails during the initial sleep. It will wake
kswapd up, however kswapd considers the zone balanced from the order-0
perspective, and will just quickly try to sleep again. So if there's a
longer stream of high-order allocations hitting the slowpath and waking
up kswapd, it might never actually wake up kcompactd, which may be
considered a regression from kswapd-based compaction. In the worst
case, it might be that a single allocation that cannot direct
reclaim/compact itself is waking kswapd in the retry loop and preventing
kcompactd from being woken up and unblocking it.
This patch makes sure kcompactd is woken up in such situations by simply
moving the wakeup before the short initial sleep. More efficient
solution would be to wake kcompactd immediately instead of kswapd if the
node is already order-0 balanced, but in that case we should also move
reset_isolation_suitable() call to kcompactd so it's not adding to the
allocator's latency. Since it's late in the 4.6 cycle, let's go with
the simpler change for now.
Fixes: accf62422b3a ("mm, kswapd: replace kswapd compaction with waking up kcompactd")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-28 23:18:49 +00:00
|
|
|
/*
|
|
|
|
* Compaction records what page blocks it recently failed to
|
|
|
|
* isolate pages from and skips them in the future scanning.
|
|
|
|
* When kswapd is going to sleep, it is reasonable to assume
|
|
|
|
* that pages and compaction may succeed so reset the cache.
|
|
|
|
*/
|
|
|
|
reset_isolation_suitable(pgdat);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We have freed the memory, now we should compact it to make
|
|
|
|
* allocation of the requested order possible.
|
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
|
mm: wake kcompactd before kswapd's short sleep
When kswapd goes to sleep it checks if the node is balanced and at first
it sleeps only for HZ/10 time, then rechecks if the node is still
balanced and nobody has woken it during the initial sleep. Only then it
goes fully sleep until an allocation slowpath wakes it up again.
For higher-order allocations, waking up kcompactd is done only before
the full sleep. This turns out to be an issue in case another
high-order allocation fails during the initial sleep. It will wake
kswapd up, however kswapd considers the zone balanced from the order-0
perspective, and will just quickly try to sleep again. So if there's a
longer stream of high-order allocations hitting the slowpath and waking
up kswapd, it might never actually wake up kcompactd, which may be
considered a regression from kswapd-based compaction. In the worst
case, it might be that a single allocation that cannot direct
reclaim/compact itself is waking kswapd in the retry loop and preventing
kcompactd from being woken up and unblocking it.
This patch makes sure kcompactd is woken up in such situations by simply
moving the wakeup before the short initial sleep. More efficient
solution would be to wake kcompactd immediately instead of kswapd if the
node is already order-0 balanced, but in that case we should also move
reset_isolation_suitable() call to kcompactd so it's not adding to the
allocator's latency. Since it's late in the 4.6 cycle, let's go with
the simpler change for now.
Fixes: accf62422b3a ("mm, kswapd: replace kswapd compaction with waking up kcompactd")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-28 23:18:49 +00:00
|
|
|
|
2011-01-13 23:45:50 +00:00
|
|
|
remaining = schedule_timeout(HZ/10);
|
2016-07-28 22:45:49 +00:00
|
|
|
|
|
|
|
/*
|
2020-06-03 22:59:01 +00:00
|
|
|
* If woken prematurely then reset kswapd_highest_zoneidx and
|
2016-07-28 22:45:49 +00:00
|
|
|
* order. The values will either be from a wakeup request or
|
|
|
|
* the previous request that slept prematurely.
|
|
|
|
*/
|
|
|
|
if (remaining) {
|
2020-06-03 22:59:01 +00:00
|
|
|
WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
|
|
|
|
kswapd_highest_zoneidx(pgdat,
|
|
|
|
highest_zoneidx));
|
2020-04-02 04:10:12 +00:00
|
|
|
|
|
|
|
if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
|
|
|
|
WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
|
2016-07-28 22:45:49 +00:00
|
|
|
}
|
|
|
|
|
2011-01-13 23:45:50 +00:00
|
|
|
finish_wait(&pgdat->kswapd_wait, &wait);
|
|
|
|
prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* After a short sleep, check if it was a premature sleep. If not, then
|
|
|
|
* go fully to sleep until explicitly woken up.
|
|
|
|
*/
|
2016-07-28 22:46:41 +00:00
|
|
|
if (!remaining &&
|
2020-06-03 22:59:01 +00:00
|
|
|
prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
|
2011-01-13 23:45:50 +00:00
|
|
|
trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* vmstat counters are not perfectly accurate and the estimated
|
|
|
|
* value for counters such as NR_FREE_PAGES can deviate from the
|
|
|
|
* true value by nr_online_cpus * threshold. To avoid the zone
|
|
|
|
* watermarks being breached while under pressure, we reduce the
|
|
|
|
* per-cpu vmstat threshold while kswapd is awake and restore
|
|
|
|
* them before going back to sleep.
|
|
|
|
*/
|
|
|
|
set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
|
2012-07-17 22:48:07 +00:00
|
|
|
|
|
|
|
if (!kthread_should_stop())
|
|
|
|
schedule();
|
|
|
|
|
2011-01-13 23:45:50 +00:00
|
|
|
set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
|
|
|
|
} else {
|
|
|
|
if (remaining)
|
|
|
|
count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
|
|
|
|
else
|
|
|
|
count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
|
|
|
|
}
|
|
|
|
finish_wait(&pgdat->kswapd_wait, &wait);
|
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* The background pageout daemon, started as a kernel thread
|
2008-10-19 03:26:32 +00:00
|
|
|
* from the init process.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* This basically trickles out pages so that we have _some_
|
|
|
|
* free memory available even if there is no other activity
|
|
|
|
* that frees anything up. This is needed for things like routing
|
|
|
|
* etc, where we otherwise might have all activity going on in
|
|
|
|
* asynchronous contexts that cannot page things out.
|
|
|
|
*
|
|
|
|
* If there are applications that are active memory-allocators
|
|
|
|
* (most normal use), this basically shouldn't matter.
|
|
|
|
*/
|
|
|
|
static int kswapd(void *p)
|
|
|
|
{
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
unsigned int alloc_order, reclaim_order;
|
2020-06-03 22:59:01 +00:00
|
|
|
unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
|
2021-05-05 01:40:12 +00:00
|
|
|
pg_data_t *pgdat = (pg_data_t *)p;
|
2005-04-16 22:20:36 +00:00
|
|
|
struct task_struct *tsk = current;
|
2009-03-13 04:19:46 +00:00
|
|
|
const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2008-12-31 23:42:29 +00:00
|
|
|
if (!cpumask_empty(cpumask))
|
2008-04-05 01:11:10 +00:00
|
|
|
set_cpus_allowed_ptr(tsk, cpumask);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Tell the memory management that we're a "memory allocator",
|
|
|
|
* and that if we need more memory we should get access to it
|
|
|
|
* regardless (see "__alloc_pages()"). "kswapd" should
|
|
|
|
* never get caught in the normal page freeing logic.
|
|
|
|
*
|
|
|
|
* (Kswapd normally doesn't need memory anyway, but sometimes
|
|
|
|
* you need a small amount of memory in order to be able to
|
|
|
|
* page out something else, and this flag essentially protects
|
|
|
|
* us from recursively trying to free more memory as we're
|
|
|
|
* trying to free the first piece of memory in the first place).
|
|
|
|
*/
|
2022-03-22 21:45:38 +00:00
|
|
|
tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
|
2007-07-17 11:03:35 +00:00
|
|
|
set_freezable();
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2020-04-02 04:10:12 +00:00
|
|
|
WRITE_ONCE(pgdat->kswapd_order, 0);
|
2020-06-03 22:59:01 +00:00
|
|
|
WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
|
mm/vmscan: throttle reclaim until some writeback completes if congested
Patch series "Remove dependency on congestion_wait in mm/", v5.
This series that removes all calls to congestion_wait in mm/ and deletes
wait_iff_congested. It's not a clever implementation but
congestion_wait has been broken for a long time [1].
Even if congestion throttling worked, it was never a great idea. While
excessive dirty/writeback pages at the tail of the LRU is one
possibility that reclaim may be slow, there is also the problem of too
many pages being isolated and reclaim failing for other reasons
(elevated references, too many pages isolated, excessive LRU contention
etc).
This series replaces the "congestion" throttling with 3 different types.
- If there are too many dirty/writeback pages, sleep until a timeout or
enough pages get cleaned
- If too many pages are isolated, sleep until enough isolated pages are
either reclaimed or put back on the LRU
- If no progress is being made, direct reclaim tasks sleep until
another task makes progress with acceptable efficiency.
This was initially tested with a mix of workloads that used to trigger
corner cases that no longer work. A new test case was created called
"stutterp" (pagereclaim-stutterp-noreaders in mmtests) using a freshly
created XFS filesystem. Note that it may be necessary to increase the
timeout of ssh if executing remotely as ssh itself can get throttled and
the connection may timeout.
stutterp varies the number of "worker" processes from 4 up to NR_CPUS*4
to check the impact as the number of direct reclaimers increase. It has
four types of worker.
- One "anon latency" worker creates small mappings with mmap() and
times how long it takes to fault the mapping reading it 4K at a time
- X file writers which is fio randomly writing X files where the total
size of the files add up to the allowed dirty_ratio. fio is allowed
to run for a warmup period to allow some file-backed pages to
accumulate. The duration of the warmup is based on the best-case
linear write speed of the storage.
- Y file readers which is fio randomly reading small files
- Z anon memory hogs which continually map (100-dirty_ratio)% of memory
- Total estimated WSS = (100+dirty_ration) percentage of memory
X+Y+Z+1 == NR_WORKERS varying from 4 up to NR_CPUS*4
The intent is to maximise the total WSS with a mix of file and anon
memory where some anonymous memory must be swapped and there is a high
likelihood of dirty/writeback pages reaching the end of the LRU.
The test can be configured to have no background readers to stress
dirty/writeback pages. The results below are based on having zero
readers.
The short summary of the results is that the series works and stalls
until some event occurs but the timeouts may need adjustment.
The test results are not broken down by patch as the series should be
treated as one block that replaces a broken throttling mechanism with a
working one.
Finally, three machines were tested but I'm reporting the worst set of
results. The other two machines had much better latencies for example.
First the results of the "anon latency" latency
stutterp
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r4
Amean mmap-4 31.4003 ( 0.00%) 2661.0198 (-8374.52%)
Amean mmap-7 38.1641 ( 0.00%) 149.2891 (-291.18%)
Amean mmap-12 60.0981 ( 0.00%) 187.8105 (-212.51%)
Amean mmap-21 161.2699 ( 0.00%) 213.9107 ( -32.64%)
Amean mmap-30 174.5589 ( 0.00%) 377.7548 (-116.41%)
Amean mmap-48 8106.8160 ( 0.00%) 1070.5616 ( 86.79%)
Stddev mmap-4 41.3455 ( 0.00%) 27573.9676 (-66591.66%)
Stddev mmap-7 53.5556 ( 0.00%) 4608.5860 (-8505.23%)
Stddev mmap-12 171.3897 ( 0.00%) 5559.4542 (-3143.75%)
Stddev mmap-21 1506.6752 ( 0.00%) 5746.2507 (-281.39%)
Stddev mmap-30 557.5806 ( 0.00%) 7678.1624 (-1277.05%)
Stddev mmap-48 61681.5718 ( 0.00%) 14507.2830 ( 76.48%)
Max-90 mmap-4 31.4243 ( 0.00%) 83.1457 (-164.59%)
Max-90 mmap-7 41.0410 ( 0.00%) 41.0720 ( -0.08%)
Max-90 mmap-12 66.5255 ( 0.00%) 53.9073 ( 18.97%)
Max-90 mmap-21 146.7479 ( 0.00%) 105.9540 ( 27.80%)
Max-90 mmap-30 193.9513 ( 0.00%) 64.3067 ( 66.84%)
Max-90 mmap-48 277.9137 ( 0.00%) 591.0594 (-112.68%)
Max mmap-4 1913.8009 ( 0.00%) 299623.9695 (-15555.96%)
Max mmap-7 2423.9665 ( 0.00%) 204453.1708 (-8334.65%)
Max mmap-12 6845.6573 ( 0.00%) 221090.3366 (-3129.64%)
Max mmap-21 56278.6508 ( 0.00%) 213877.3496 (-280.03%)
Max mmap-30 19716.2990 ( 0.00%) 216287.6229 (-997.00%)
Max mmap-48 477923.9400 ( 0.00%) 245414.8238 ( 48.65%)
For most thread counts, the time to mmap() is unfortunately increased.
In earlier versions of the series, this was lower but a large number of
throttling events were reaching their timeout increasing the amount of
inefficient scanning of the LRU. There is no prioritisation of reclaim
tasks making progress based on each tasks rate of page allocation versus
progress of reclaim. The variance is also impacted for high worker
counts but in all cases, the differences in latency are not
statistically significant due to very large maximum outliers. Max-90
shows that 90% of the stalls are comparable but the Max results show the
massive outliers which are increased to to stalling.
It is expected that this will be very machine dependant. Due to the
test design, reclaim is difficult so allocations stall and there are
variances depending on whether THPs can be allocated or not. The amount
of memory will affect exactly how bad the corner cases are and how often
they trigger. The warmup period calculation is not ideal as it's based
on linear writes where as fio is randomly writing multiple files from
multiple tasks so the start state of the test is variable. For example,
these are the latencies on a single-socket machine that had more memory
Amean mmap-4 42.2287 ( 0.00%) 49.6838 * -17.65%*
Amean mmap-7 216.4326 ( 0.00%) 47.4451 * 78.08%*
Amean mmap-12 2412.0588 ( 0.00%) 51.7497 ( 97.85%)
Amean mmap-21 5546.2548 ( 0.00%) 51.8862 ( 99.06%)
Amean mmap-30 1085.3121 ( 0.00%) 72.1004 ( 93.36%)
The overall system CPU usage and elapsed time is as follows
5.15.0-rc3 5.15.0-rc3
vanilla mm-reclaimcongest-v5r4
Duration User 6989.03 983.42
Duration System 7308.12 799.68
Duration Elapsed 2277.67 2092.98
The patches reduce system CPU usage by 89% as the vanilla kernel is rarely
stalling.
The high-level /proc/vmstats show
5.15.0-rc1 5.15.0-rc1
vanilla mm-reclaimcongest-v5r2
Ops Direct pages scanned 1056608451.00 503594991.00
Ops Kswapd pages scanned 109795048.00 147289810.00
Ops Kswapd pages reclaimed 63269243.00 31036005.00
Ops Direct pages reclaimed 10803973.00 6328887.00
Ops Kswapd efficiency % 57.62 21.07
Ops Kswapd velocity 48204.98 57572.86
Ops Direct efficiency % 1.02 1.26
Ops Direct velocity 463898.83 196845.97
Kswapd scanned less pages but the detailed pattern is different. The
vanilla kernel scans slowly over time where as the patches exhibits
burst patterns of scan activity. Direct reclaim scanning is reduced by
52% due to stalling.
The pattern for stealing pages is also slightly different. Both kernels
exhibit spikes but the vanilla kernel when reclaiming shows pages being
reclaimed over a period of time where as the patches tend to reclaim in
spikes. The difference is that vanilla is not throttling and instead
scanning constantly finding some pages over time where as the patched
kernel throttles and reclaims in spikes.
Ops Percentage direct scans 90.59 77.37
For direct reclaim, vanilla scanned 90.59% of pages where as with the
patches, 77.37% were direct reclaim due to throttling
Ops Page writes by reclaim 2613590.00 1687131.00
Page writes from reclaim context are reduced.
Ops Page writes anon 2932752.00 1917048.00
And there is less swapping.
Ops Page reclaim immediate 996248528.00 107664764.00
The number of pages encountered at the tail of the LRU tagged for
immediate reclaim but still dirty/writeback is reduced by 89%.
Ops Slabs scanned 164284.00 153608.00
Slab scan activity is similar.
ftrace was used to gather stall activity
Vanilla
-------
1 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=16000
2 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=12000
8 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=8000
29 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=4000
82394 writeback_wait_iff_congested: usec_timeout=100000 usec_delayed=0
The fast majority of wait_iff_congested calls do not stall at all. What
is likely happening is that cond_resched() reschedules the task for a
short period when the BDI is not registering congestion (which it never
will in this test setup).
1 writeback_congestion_wait: usec_timeout=100000 usec_delayed=120000
2 writeback_congestion_wait: usec_timeout=100000 usec_delayed=132000
4 writeback_congestion_wait: usec_timeout=100000 usec_delayed=112000
380 writeback_congestion_wait: usec_timeout=100000 usec_delayed=108000
778 writeback_congestion_wait: usec_timeout=100000 usec_delayed=104000
congestion_wait if called always exceeds the timeout as there is no
trigger to wake it up.
Bottom line: Vanilla will throttle but it's not effective.
Patch series
------------
Kswapd throttle activity was always due to scanning pages tagged for
immediate reclaim at the tail of the LRU
1 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
4 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
94 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
112 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority of events did not stall or stalled for a short period.
Roughly 16% of stalls reached the timeout before expiry. For direct
reclaim, the number of times stalled for each reason were
6624 reason=VMSCAN_THROTTLE_ISOLATED
93246 reason=VMSCAN_THROTTLE_NOPROGRESS
96934 reason=VMSCAN_THROTTLE_WRITEBACK
The most common reason to stall was due to excessive pages tagged for
immediate reclaim at the tail of the LRU followed by a failure to make
forward. A relatively small number were due to too many pages isolated
from the LRU by parallel threads
For VMSCAN_THROTTLE_ISOLATED, the breakdown of delays was
9 usec_timeout=20000 usect_delayed=4000 reason=VMSCAN_THROTTLE_ISOLATED
12 usec_timeout=20000 usect_delayed=16000 reason=VMSCAN_THROTTLE_ISOLATED
83 usec_timeout=20000 usect_delayed=20000 reason=VMSCAN_THROTTLE_ISOLATED
6520 usec_timeout=20000 usect_delayed=0 reason=VMSCAN_THROTTLE_ISOLATED
Most did not stall at all. A small number reached the timeout.
For VMSCAN_THROTTLE_NOPROGRESS, the breakdown of stalls were all over
the map
1 usec_timeout=500000 usect_delayed=324000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=332000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=348000 reason=VMSCAN_THROTTLE_NOPROGRESS
1 usec_timeout=500000 usect_delayed=360000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=228000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=260000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=340000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=364000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=372000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=428000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=460000 reason=VMSCAN_THROTTLE_NOPROGRESS
2 usec_timeout=500000 usect_delayed=464000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=244000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=252000 reason=VMSCAN_THROTTLE_NOPROGRESS
3 usec_timeout=500000 usect_delayed=272000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=188000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=268000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=328000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=380000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=392000 reason=VMSCAN_THROTTLE_NOPROGRESS
4 usec_timeout=500000 usect_delayed=432000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=204000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=220000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=412000 reason=VMSCAN_THROTTLE_NOPROGRESS
5 usec_timeout=500000 usect_delayed=436000 reason=VMSCAN_THROTTLE_NOPROGRESS
6 usec_timeout=500000 usect_delayed=488000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=212000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=300000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=316000 reason=VMSCAN_THROTTLE_NOPROGRESS
7 usec_timeout=500000 usect_delayed=472000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=248000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=356000 reason=VMSCAN_THROTTLE_NOPROGRESS
8 usec_timeout=500000 usect_delayed=456000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=124000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=376000 reason=VMSCAN_THROTTLE_NOPROGRESS
9 usec_timeout=500000 usect_delayed=484000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=172000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=420000 reason=VMSCAN_THROTTLE_NOPROGRESS
10 usec_timeout=500000 usect_delayed=452000 reason=VMSCAN_THROTTLE_NOPROGRESS
11 usec_timeout=500000 usect_delayed=256000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=112000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=116000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=144000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=152000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=264000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=384000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=424000 reason=VMSCAN_THROTTLE_NOPROGRESS
12 usec_timeout=500000 usect_delayed=492000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=184000 reason=VMSCAN_THROTTLE_NOPROGRESS
13 usec_timeout=500000 usect_delayed=444000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=308000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=440000 reason=VMSCAN_THROTTLE_NOPROGRESS
14 usec_timeout=500000 usect_delayed=476000 reason=VMSCAN_THROTTLE_NOPROGRESS
16 usec_timeout=500000 usect_delayed=140000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=232000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=240000 reason=VMSCAN_THROTTLE_NOPROGRESS
17 usec_timeout=500000 usect_delayed=280000 reason=VMSCAN_THROTTLE_NOPROGRESS
18 usec_timeout=500000 usect_delayed=404000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=148000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=216000 reason=VMSCAN_THROTTLE_NOPROGRESS
20 usec_timeout=500000 usect_delayed=468000 reason=VMSCAN_THROTTLE_NOPROGRESS
21 usec_timeout=500000 usect_delayed=448000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=168000 reason=VMSCAN_THROTTLE_NOPROGRESS
23 usec_timeout=500000 usect_delayed=296000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=132000 reason=VMSCAN_THROTTLE_NOPROGRESS
25 usec_timeout=500000 usect_delayed=352000 reason=VMSCAN_THROTTLE_NOPROGRESS
26 usec_timeout=500000 usect_delayed=180000 reason=VMSCAN_THROTTLE_NOPROGRESS
27 usec_timeout=500000 usect_delayed=284000 reason=VMSCAN_THROTTLE_NOPROGRESS
28 usec_timeout=500000 usect_delayed=164000 reason=VMSCAN_THROTTLE_NOPROGRESS
29 usec_timeout=500000 usect_delayed=136000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=200000 reason=VMSCAN_THROTTLE_NOPROGRESS
30 usec_timeout=500000 usect_delayed=400000 reason=VMSCAN_THROTTLE_NOPROGRESS
31 usec_timeout=500000 usect_delayed=196000 reason=VMSCAN_THROTTLE_NOPROGRESS
32 usec_timeout=500000 usect_delayed=156000 reason=VMSCAN_THROTTLE_NOPROGRESS
33 usec_timeout=500000 usect_delayed=224000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=128000 reason=VMSCAN_THROTTLE_NOPROGRESS
35 usec_timeout=500000 usect_delayed=176000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=368000 reason=VMSCAN_THROTTLE_NOPROGRESS
36 usec_timeout=500000 usect_delayed=496000 reason=VMSCAN_THROTTLE_NOPROGRESS
37 usec_timeout=500000 usect_delayed=312000 reason=VMSCAN_THROTTLE_NOPROGRESS
38 usec_timeout=500000 usect_delayed=304000 reason=VMSCAN_THROTTLE_NOPROGRESS
40 usec_timeout=500000 usect_delayed=288000 reason=VMSCAN_THROTTLE_NOPROGRESS
43 usec_timeout=500000 usect_delayed=408000 reason=VMSCAN_THROTTLE_NOPROGRESS
55 usec_timeout=500000 usect_delayed=416000 reason=VMSCAN_THROTTLE_NOPROGRESS
56 usec_timeout=500000 usect_delayed=76000 reason=VMSCAN_THROTTLE_NOPROGRESS
58 usec_timeout=500000 usect_delayed=120000 reason=VMSCAN_THROTTLE_NOPROGRESS
59 usec_timeout=500000 usect_delayed=208000 reason=VMSCAN_THROTTLE_NOPROGRESS
61 usec_timeout=500000 usect_delayed=68000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=192000 reason=VMSCAN_THROTTLE_NOPROGRESS
71 usec_timeout=500000 usect_delayed=480000 reason=VMSCAN_THROTTLE_NOPROGRESS
79 usec_timeout=500000 usect_delayed=60000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=320000 reason=VMSCAN_THROTTLE_NOPROGRESS
82 usec_timeout=500000 usect_delayed=92000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=64000 reason=VMSCAN_THROTTLE_NOPROGRESS
85 usec_timeout=500000 usect_delayed=80000 reason=VMSCAN_THROTTLE_NOPROGRESS
88 usec_timeout=500000 usect_delayed=84000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=160000 reason=VMSCAN_THROTTLE_NOPROGRESS
90 usec_timeout=500000 usect_delayed=292000 reason=VMSCAN_THROTTLE_NOPROGRESS
94 usec_timeout=500000 usect_delayed=56000 reason=VMSCAN_THROTTLE_NOPROGRESS
118 usec_timeout=500000 usect_delayed=88000 reason=VMSCAN_THROTTLE_NOPROGRESS
119 usec_timeout=500000 usect_delayed=72000 reason=VMSCAN_THROTTLE_NOPROGRESS
126 usec_timeout=500000 usect_delayed=108000 reason=VMSCAN_THROTTLE_NOPROGRESS
146 usec_timeout=500000 usect_delayed=52000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=36000 reason=VMSCAN_THROTTLE_NOPROGRESS
148 usec_timeout=500000 usect_delayed=48000 reason=VMSCAN_THROTTLE_NOPROGRESS
159 usec_timeout=500000 usect_delayed=28000 reason=VMSCAN_THROTTLE_NOPROGRESS
178 usec_timeout=500000 usect_delayed=44000 reason=VMSCAN_THROTTLE_NOPROGRESS
183 usec_timeout=500000 usect_delayed=40000 reason=VMSCAN_THROTTLE_NOPROGRESS
237 usec_timeout=500000 usect_delayed=100000 reason=VMSCAN_THROTTLE_NOPROGRESS
266 usec_timeout=500000 usect_delayed=32000 reason=VMSCAN_THROTTLE_NOPROGRESS
313 usec_timeout=500000 usect_delayed=24000 reason=VMSCAN_THROTTLE_NOPROGRESS
347 usec_timeout=500000 usect_delayed=96000 reason=VMSCAN_THROTTLE_NOPROGRESS
470 usec_timeout=500000 usect_delayed=20000 reason=VMSCAN_THROTTLE_NOPROGRESS
559 usec_timeout=500000 usect_delayed=16000 reason=VMSCAN_THROTTLE_NOPROGRESS
964 usec_timeout=500000 usect_delayed=12000 reason=VMSCAN_THROTTLE_NOPROGRESS
2001 usec_timeout=500000 usect_delayed=104000 reason=VMSCAN_THROTTLE_NOPROGRESS
2447 usec_timeout=500000 usect_delayed=8000 reason=VMSCAN_THROTTLE_NOPROGRESS
7888 usec_timeout=500000 usect_delayed=4000 reason=VMSCAN_THROTTLE_NOPROGRESS
22727 usec_timeout=500000 usect_delayed=0 reason=VMSCAN_THROTTLE_NOPROGRESS
51305 usec_timeout=500000 usect_delayed=500000 reason=VMSCAN_THROTTLE_NOPROGRESS
The full timeout is often hit but a large number also do not stall at
all. The remainder slept a little allowing other reclaim tasks to make
progress.
While this timeout could be further increased, it could also negatively
impact worst-case behaviour when there is no prioritisation of what task
should make progress.
For VMSCAN_THROTTLE_WRITEBACK, the breakdown was
1 usec_timeout=100000 usect_delayed=44000 reason=VMSCAN_THROTTLE_WRITEBACK
2 usec_timeout=100000 usect_delayed=76000 reason=VMSCAN_THROTTLE_WRITEBACK
3 usec_timeout=100000 usect_delayed=80000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=48000 reason=VMSCAN_THROTTLE_WRITEBACK
5 usec_timeout=100000 usect_delayed=84000 reason=VMSCAN_THROTTLE_WRITEBACK
6 usec_timeout=100000 usect_delayed=72000 reason=VMSCAN_THROTTLE_WRITEBACK
7 usec_timeout=100000 usect_delayed=88000 reason=VMSCAN_THROTTLE_WRITEBACK
11 usec_timeout=100000 usect_delayed=56000 reason=VMSCAN_THROTTLE_WRITEBACK
12 usec_timeout=100000 usect_delayed=64000 reason=VMSCAN_THROTTLE_WRITEBACK
16 usec_timeout=100000 usect_delayed=92000 reason=VMSCAN_THROTTLE_WRITEBACK
24 usec_timeout=100000 usect_delayed=68000 reason=VMSCAN_THROTTLE_WRITEBACK
28 usec_timeout=100000 usect_delayed=32000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=60000 reason=VMSCAN_THROTTLE_WRITEBACK
30 usec_timeout=100000 usect_delayed=96000 reason=VMSCAN_THROTTLE_WRITEBACK
32 usec_timeout=100000 usect_delayed=52000 reason=VMSCAN_THROTTLE_WRITEBACK
42 usec_timeout=100000 usect_delayed=40000 reason=VMSCAN_THROTTLE_WRITEBACK
77 usec_timeout=100000 usect_delayed=28000 reason=VMSCAN_THROTTLE_WRITEBACK
99 usec_timeout=100000 usect_delayed=36000 reason=VMSCAN_THROTTLE_WRITEBACK
137 usec_timeout=100000 usect_delayed=24000 reason=VMSCAN_THROTTLE_WRITEBACK
190 usec_timeout=100000 usect_delayed=20000 reason=VMSCAN_THROTTLE_WRITEBACK
339 usec_timeout=100000 usect_delayed=16000 reason=VMSCAN_THROTTLE_WRITEBACK
518 usec_timeout=100000 usect_delayed=12000 reason=VMSCAN_THROTTLE_WRITEBACK
852 usec_timeout=100000 usect_delayed=8000 reason=VMSCAN_THROTTLE_WRITEBACK
3359 usec_timeout=100000 usect_delayed=4000 reason=VMSCAN_THROTTLE_WRITEBACK
7147 usec_timeout=100000 usect_delayed=0 reason=VMSCAN_THROTTLE_WRITEBACK
83962 usec_timeout=100000 usect_delayed=100000 reason=VMSCAN_THROTTLE_WRITEBACK
The majority hit the timeout in direct reclaim context although a
sizable number did not stall at all. This is very different to kswapd
where only a tiny percentage of stalls due to writeback reached the
timeout.
Bottom line, the throttling appears to work and the wakeup events may
limit worst case stalls. There might be some grounds for adjusting
timeouts but it's likely futile as the worst-case scenarios depend on
the workload, memory size and the speed of the storage. A better
approach to improve the series further would be to prioritise tasks
based on their rate of allocation with the caveat that it may be very
expensive to track.
This patch (of 5):
Page reclaim throttles on wait_iff_congested under the following
conditions:
- kswapd is encountering pages under writeback and marked for immediate
reclaim implying that pages are cycling through the LRU faster than
pages can be cleaned.
- Direct reclaim will stall if all dirty pages are backed by congested
inodes.
wait_iff_congested is almost completely broken with few exceptions.
This patch adds a new node-based workqueue and tracks the number of
throttled tasks and pages written back since throttling started. If
enough pages belonging to the node are written back then the throttled
tasks will wake early. If not, the throttled tasks sleeps until the
timeout expires.
[neilb@suse.de: Uninterruptible sleep and simpler wakeups]
[hdanton@sina.com: Avoid race when reclaim starts]
[vbabka@suse.cz: vmstat irq-safe api, clarifications]
Link: https://lore.kernel.org/linux-mm/45d8b7a6-8548-65f5-cccf-9f451d4ae3d4@kernel.dk/ [1]
Link: https://lkml.kernel.org/r/20211022144651.19914-1-mgorman@techsingularity.net
Link: https://lkml.kernel.org/r/20211022144651.19914-2-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: NeilBrown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Cc: "Darrick J . Wong" <djwong@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:42:25 +00:00
|
|
|
atomic_set(&pgdat->nr_writeback_throttled, 0);
|
2005-04-16 22:20:36 +00:00
|
|
|
for ( ; ; ) {
|
2012-12-12 00:02:48 +00:00
|
|
|
bool ret;
|
2005-06-25 06:13:50 +00:00
|
|
|
|
2020-04-02 04:10:12 +00:00
|
|
|
alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
|
2020-06-03 22:59:01 +00:00
|
|
|
highest_zoneidx = kswapd_highest_zoneidx(pgdat,
|
|
|
|
highest_zoneidx);
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
|
2016-07-28 22:45:49 +00:00
|
|
|
kswapd_try_sleep:
|
|
|
|
kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
|
2020-06-03 22:59:01 +00:00
|
|
|
highest_zoneidx);
|
mm: vmscan: only read new_classzone_idx from pgdat when reclaiming successfully
During allocator-intensive workloads, kswapd will be woken frequently
causing free memory to oscillate between the high and min watermark. This
is expected behaviour. Unfortunately, if the highest zone is small, a
problem occurs.
When balance_pgdat() returns, it may be at a lower classzone_idx than it
started because the highest zone was unreclaimable. Before checking if it
should go to sleep though, it checks pgdat->classzone_idx which when there
is no other activity will be MAX_NR_ZONES-1. It interprets this as it has
been woken up while reclaiming, skips scheduling and reclaims again. As
there is no useful reclaim work to do, it enters into a loop of shrinking
slab consuming loads of CPU until the highest zone becomes reclaimable for
a long period of time.
There are two problems here. 1) If the returned classzone or order is
lower, it'll continue reclaiming without scheduling. 2) if the highest
zone was marked unreclaimable but balance_pgdat() returns immediately at
DEF_PRIORITY, the new lower classzone is not communicated back to kswapd()
for sleeping.
This patch does two things that are related. If the end_zone is
unreclaimable, this information is communicated back. Second, if the
classzone or order was reduced due to failing to reclaim, new information
is not read from pgdat and instead an attempt is made to go to sleep. Due
to this, it is also necessary that pgdat->classzone_idx be initialised
each time to pgdat->nr_zones - 1 to avoid re-reads being interpreted as
wakeups.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reported-by: Pádraig Brady <P@draigBrady.com>
Tested-by: Pádraig Brady <P@draigBrady.com>
Tested-by: Andrew Lutomirski <luto@mit.edu>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-08 22:39:40 +00:00
|
|
|
|
2020-06-03 22:59:01 +00:00
|
|
|
/* Read the new order and highest_zoneidx */
|
2020-12-15 03:12:18 +00:00
|
|
|
alloc_order = READ_ONCE(pgdat->kswapd_order);
|
2020-06-03 22:59:01 +00:00
|
|
|
highest_zoneidx = kswapd_highest_zoneidx(pgdat,
|
|
|
|
highest_zoneidx);
|
2020-04-02 04:10:12 +00:00
|
|
|
WRITE_ONCE(pgdat->kswapd_order, 0);
|
2020-06-03 22:59:01 +00:00
|
|
|
WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2009-12-15 01:58:33 +00:00
|
|
|
ret = try_to_freeze();
|
|
|
|
if (kthread_should_stop())
|
|
|
|
break;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We can speed up thawing tasks if we don't call balance_pgdat
|
|
|
|
* after returning from the refrigerator
|
|
|
|
*/
|
2016-07-28 22:45:49 +00:00
|
|
|
if (ret)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Reclaim begins at the requested order but if a high-order
|
|
|
|
* reclaim fails then kswapd falls back to reclaiming for
|
|
|
|
* order-0. If that happens, kswapd will consider sleeping
|
|
|
|
* for the order it finished reclaiming at (reclaim_order)
|
|
|
|
* but kcompactd is woken to compact for the original
|
|
|
|
* request (alloc_order).
|
|
|
|
*/
|
2020-06-03 22:59:01 +00:00
|
|
|
trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
|
2016-07-28 22:46:47 +00:00
|
|
|
alloc_order);
|
2020-06-03 22:59:01 +00:00
|
|
|
reclaim_order = balance_pgdat(pgdat, alloc_order,
|
|
|
|
highest_zoneidx);
|
2016-07-28 22:45:49 +00:00
|
|
|
if (reclaim_order < alloc_order)
|
|
|
|
goto kswapd_try_sleep;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2012-11-08 23:53:39 +00:00
|
|
|
|
2022-03-22 21:45:38 +00:00
|
|
|
tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
|
mm: vmscan: clear kswapd's special reclaim powers before exiting
When kswapd exits, it can end up taking locks that were previously held
by allocating tasks while they waited for reclaim. Lockdep currently
warns about this:
On Wed, May 28, 2014 at 06:06:34PM +0800, Gu Zheng wrote:
> inconsistent {RECLAIM_FS-ON-W} -> {IN-RECLAIM_FS-R} usage.
> kswapd2/1151 [HC0[0]:SC0[0]:HE1:SE1] takes:
> (&sig->group_rwsem){+++++?}, at: exit_signals+0x24/0x130
> {RECLAIM_FS-ON-W} state was registered at:
> mark_held_locks+0xb9/0x140
> lockdep_trace_alloc+0x7a/0xe0
> kmem_cache_alloc_trace+0x37/0x240
> flex_array_alloc+0x99/0x1a0
> cgroup_attach_task+0x63/0x430
> attach_task_by_pid+0x210/0x280
> cgroup_procs_write+0x16/0x20
> cgroup_file_write+0x120/0x2c0
> vfs_write+0xc0/0x1f0
> SyS_write+0x4c/0xa0
> tracesys+0xdd/0xe2
> irq event stamp: 49
> hardirqs last enabled at (49): _raw_spin_unlock_irqrestore+0x36/0x70
> hardirqs last disabled at (48): _raw_spin_lock_irqsave+0x2b/0xa0
> softirqs last enabled at (0): copy_process.part.24+0x627/0x15f0
> softirqs last disabled at (0): (null)
>
> other info that might help us debug this:
> Possible unsafe locking scenario:
>
> CPU0
> ----
> lock(&sig->group_rwsem);
> <Interrupt>
> lock(&sig->group_rwsem);
>
> *** DEADLOCK ***
>
> no locks held by kswapd2/1151.
>
> stack backtrace:
> CPU: 30 PID: 1151 Comm: kswapd2 Not tainted 3.10.39+ #4
> Call Trace:
> dump_stack+0x19/0x1b
> print_usage_bug+0x1f7/0x208
> mark_lock+0x21d/0x2a0
> __lock_acquire+0x52a/0xb60
> lock_acquire+0xa2/0x140
> down_read+0x51/0xa0
> exit_signals+0x24/0x130
> do_exit+0xb5/0xa50
> kthread+0xdb/0x100
> ret_from_fork+0x7c/0xb0
This is because the kswapd thread is still marked as a reclaimer at the
time of exit. But because it is exiting, nobody is actually waiting on
it to make reclaim progress anymore, and it's nothing but a regular
thread at this point. Be tidy and strip it of all its powers
(PF_MEMALLOC, PF_SWAPWRITE, PF_KSWAPD, and the lockdep reclaim state)
before returning from the thread function.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Tang Chen <tangchen@cn.fujitsu.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-06 21:35:35 +00:00
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2018-04-05 23:25:16 +00:00
|
|
|
* A zone is low on free memory or too fragmented for high-order memory. If
|
|
|
|
* kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
|
|
|
|
* pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim
|
|
|
|
* has failed or is not needed, still wake up kcompactd if only compaction is
|
|
|
|
* needed.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2018-04-05 23:25:16 +00:00
|
|
|
void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
|
2020-06-03 22:59:01 +00:00
|
|
|
enum zone_type highest_zoneidx)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
|
|
|
pg_data_t *pgdat;
|
2020-04-02 04:10:12 +00:00
|
|
|
enum zone_type curr_idx;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2016-09-01 23:14:55 +00:00
|
|
|
if (!managed_zone(zone))
|
2005-04-16 22:20:36 +00:00
|
|
|
return;
|
|
|
|
|
2018-04-05 23:25:16 +00:00
|
|
|
if (!cpuset_zone_allowed(zone, gfp_flags))
|
2005-04-16 22:20:36 +00:00
|
|
|
return;
|
2020-04-02 04:10:12 +00:00
|
|
|
|
mm: page allocator: adjust the per-cpu counter threshold when memory is low
Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory
is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To
avoid synchronization overhead, these counters are maintained on a per-cpu
basis and drained both periodically and when a threshold is above a
threshold. On large CPU systems, the difference between the estimate and
real value of NR_FREE_PAGES can be very high. The system can get into a
case where pages are allocated far below the min watermark potentially
causing livelock issues. The commit solved the problem by taking a better
reading of NR_FREE_PAGES when memory was low.
Unfortately, as reported by Shaohua Li this accurate reading can consume a
large amount of CPU time on systems with many sockets due to cache line
bouncing. This patch takes a different approach. For large machines
where counter drift might be unsafe and while kswapd is awake, the per-cpu
thresholds for the target pgdat are reduced to limit the level of drift to
what should be a safe level. This incurs a performance penalty in heavy
memory pressure by a factor that depends on the workload and the machine
but the machine should function correctly without accidentally exhausting
all memory on a node. There is an additional cost when kswapd wakes and
sleeps but the event is not expected to be frequent - in Shaohua's test
case, there was one recorded sleep and wake event at least.
To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is
introduced that takes a more accurate reading of NR_FREE_PAGES when called
from wakeup_kswapd, when deciding whether it is really safe to go back to
sleep in sleeping_prematurely() and when deciding if a zone is really
balanced or not in balance_pgdat(). We are still using an expensive
function but limiting how often it is called.
When the test case is reproduced, the time spent in the watermark
functions is reduced. The following report is on the percentage of time
spent cumulatively spent in the functions zone_nr_free_pages(),
zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(),
zone_page_state_snapshot(), zone_page_state().
vanilla 11.6615%
disable-threshold 0.2584%
David said:
: We had to pull aa454840 "mm: page allocator: calculate a better estimate
: of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36
: internally because tests showed that it would cause the machine to stall
: as the result of heavy kswapd activity. I merged it back with this fix as
: it is pending in the -mm tree and it solves the issue we were seeing, so I
: definitely think this should be pushed to -stable (and I would seriously
: consider it for 2.6.37 inclusion even at this late date).
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reported-by: Shaohua Li <shaohua.li@intel.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Tested-by: Nicolas Bareil <nico@chdir.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
|
|
|
pgdat = zone->zone_pgdat;
|
2020-06-03 22:59:01 +00:00
|
|
|
curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
|
2020-04-02 04:10:12 +00:00
|
|
|
|
2020-06-03 22:59:01 +00:00
|
|
|
if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
|
|
|
|
WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
|
2020-04-02 04:10:12 +00:00
|
|
|
|
|
|
|
if (READ_ONCE(pgdat->kswapd_order) < order)
|
|
|
|
WRITE_ONCE(pgdat->kswapd_order, order);
|
2019-07-04 22:14:42 +00:00
|
|
|
|
2005-09-13 08:25:07 +00:00
|
|
|
if (!waitqueue_active(&pgdat->kswapd_wait))
|
2005-04-16 22:20:36 +00:00
|
|
|
return;
|
2016-07-28 22:46:26 +00:00
|
|
|
|
2018-04-05 23:25:16 +00:00
|
|
|
/* Hopeless node, leave it to direct reclaim if possible */
|
|
|
|
if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
|
2020-06-03 22:59:01 +00:00
|
|
|
(pgdat_balanced(pgdat, order, highest_zoneidx) &&
|
|
|
|
!pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
|
2018-04-05 23:25:16 +00:00
|
|
|
/*
|
|
|
|
* There may be plenty of free memory available, but it's too
|
|
|
|
* fragmented for high-order allocations. Wake up kcompactd
|
|
|
|
* and rely on compaction_suitable() to determine if it's
|
|
|
|
* needed. If it fails, it will defer subsequent attempts to
|
|
|
|
* ratelimit its work.
|
|
|
|
*/
|
|
|
|
if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
|
2020-06-03 22:59:01 +00:00
|
|
|
wakeup_kcompactd(pgdat, order, highest_zoneidx);
|
mm, vmscan: prevent kswapd sleeping prematurely due to mismatched classzone_idx
kswapd is woken to reclaim a node based on a failed allocation request
from any eligible zone. Once reclaiming in balance_pgdat(), it will
continue reclaiming until there is an eligible zone available for the
zone it was woken for. kswapd tracks what zone it was recently woken
for in pgdat->kswapd_classzone_idx. If it has not been woken recently,
this zone will be 0.
However, the decision on whether to sleep is made on
kswapd_classzone_idx which is 0 without a recent wakeup request and that
classzone does not account for lowmem reserves. This allows kswapd to
sleep when a low small zone such as ZONE_DMA is balanced for a GFP_DMA
request even if a stream of allocations cannot use that zone. While
kswapd may be woken again shortly in the near future there are two
consequences -- the pgdat bits that control congestion are cleared
prematurely and direct reclaim is more likely as kswapd slept
prematurely.
This patch flips kswapd_classzone_idx to default to MAX_NR_ZONES (an
invalid index) when there has been no recent wakeups. If there are no
wakeups, it'll decide whether to sleep based on the highest possible
zone available (MAX_NR_ZONES - 1). It then becomes critical that the
"pgdat balanced" decisions during reclaim and when deciding to sleep are
the same. If there is a mismatch, kswapd can stay awake continually
trying to balance tiny zones.
simoop was used to evaluate it again. Two of the preparation patches
regressed the workload so they are included as the second set of
results. Otherwise this patch looks artifically excellent
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 19786774.76 ( 8.69%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 24101956.27 ( 5.32%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 27691872.71 ( 5.71%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 1011.91 ( 27.22%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 34874.98 ( 91.55%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 575449.60 ( 91.37%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 84246.26 ( -7.03%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 400058.43 (-127.91%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 10905600.00 (-4315.17%) 125765.57 ( 49.08%)
With this patch on top, write and allocation latencies are massively
improved. The read latencies are slightly impaired but it's worth
noting that this is mostly due to the IO scheduler and not directly
related to reclaim. The vmstats are a bit of a mix but the relevant
ones are as follows;
4.10.0-rc7 4.10.0-rc7 4.10.0-rc7
mmots-20170209 clear-v1r25keepawake-v1r25
Swap Ins 0 0 0
Swap Outs 0 608 0
Direct pages scanned 6910672 3132699 6357298
Kswapd pages scanned 57036946 82488665 56986286
Kswapd pages reclaimed 55993488 63474329 55939113
Direct pages reclaimed 6905990 2964843 6352115
Kswapd efficiency 98% 76% 98%
Kswapd velocity 12494.375 17597.507 12488.065
Direct efficiency 99% 94% 99%
Direct velocity 1513.835 668.306 1393.148
Page writes by reclaim 0.000 4410243.000 0.000
Page writes file 0 4409635 0
Page writes anon 0 608 0
Page reclaim immediate 1036792 14175203 1042571
4.11.0-rc1 4.11.0-rc1 4.11.0-rc1
vanilla clear-v2 keepawake-v2
Swap Ins 0 12 0
Swap Outs 0 838 0
Direct pages scanned 6579706 3237270 6256811
Kswapd pages scanned 61853702 79961486 54837791
Kswapd pages reclaimed 60768764 60755788 53849586
Direct pages reclaimed 6579055 2987453 6256151
Kswapd efficiency 98% 75% 98%
Page writes by reclaim 0.000 4389496.000 0.000
Page writes file 0 4388658 0
Page writes anon 0 838 0
Page reclaim immediate 1073573 14473009 982507
Swap-outs are equivalent to baseline.
Direct reclaim is reduced but not eliminated. It's worth noting that
there are two periods of direct reclaim for this workload. The first is
when it switches from preparing the files for the actual test itself.
It's a lot of file IO followed by a lot of allocs that reclaims heavily
for a brief window. While direct reclaim is lower with clear-v2, it is
due to kswapd scanning aggressively and trying to reclaim the world
which is not the right thing to do. With the patches applied, there is
still direct reclaim but the phase change from "creating work files" to
starting multiple threads that allocate a lot of anonymous memory faster
than kswapd can reclaim.
Scanning/reclaim efficiency is restored by this patch.
Page writes from reclaim context are back at 0 which is ideal.
Pages immediately reclaimed after IO completes is slightly improved but
it is expected this will vary slightly.
On UMA, there is almost no change so this is not expected to be a
universal win.
[mgorman@suse.de: fix ->kswapd_classzone_idx initialization]
Link: http://lkml.kernel.org/r/20170406174538.5msrznj6nt6qpbx5@suse.de
Link: http://lkml.kernel.org/r/20170309075657.25121-4-mgorman@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Shantanu Goel <sgoel01@yahoo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:53:45 +00:00
|
|
|
return;
|
2018-04-05 23:25:16 +00:00
|
|
|
}
|
mm: page allocator: adjust the per-cpu counter threshold when memory is low
Commit aa45484 ("calculate a better estimate of NR_FREE_PAGES when memory
is low") noted that watermarks were based on the vmstat NR_FREE_PAGES. To
avoid synchronization overhead, these counters are maintained on a per-cpu
basis and drained both periodically and when a threshold is above a
threshold. On large CPU systems, the difference between the estimate and
real value of NR_FREE_PAGES can be very high. The system can get into a
case where pages are allocated far below the min watermark potentially
causing livelock issues. The commit solved the problem by taking a better
reading of NR_FREE_PAGES when memory was low.
Unfortately, as reported by Shaohua Li this accurate reading can consume a
large amount of CPU time on systems with many sockets due to cache line
bouncing. This patch takes a different approach. For large machines
where counter drift might be unsafe and while kswapd is awake, the per-cpu
thresholds for the target pgdat are reduced to limit the level of drift to
what should be a safe level. This incurs a performance penalty in heavy
memory pressure by a factor that depends on the workload and the machine
but the machine should function correctly without accidentally exhausting
all memory on a node. There is an additional cost when kswapd wakes and
sleeps but the event is not expected to be frequent - in Shaohua's test
case, there was one recorded sleep and wake event at least.
To ensure that kswapd wakes up, a safe version of zone_watermark_ok() is
introduced that takes a more accurate reading of NR_FREE_PAGES when called
from wakeup_kswapd, when deciding whether it is really safe to go back to
sleep in sleeping_prematurely() and when deciding if a zone is really
balanced or not in balance_pgdat(). We are still using an expensive
function but limiting how often it is called.
When the test case is reproduced, the time spent in the watermark
functions is reduced. The following report is on the percentage of time
spent cumulatively spent in the functions zone_nr_free_pages(),
zone_watermark_ok(), __zone_watermark_ok(), zone_watermark_ok_safe(),
zone_page_state_snapshot(), zone_page_state().
vanilla 11.6615%
disable-threshold 0.2584%
David said:
: We had to pull aa454840 "mm: page allocator: calculate a better estimate
: of NR_FREE_PAGES when memory is low and kswapd is awake" from 2.6.36
: internally because tests showed that it would cause the machine to stall
: as the result of heavy kswapd activity. I merged it back with this fix as
: it is pending in the -mm tree and it solves the issue we were seeing, so I
: definitely think this should be pushed to -stable (and I would seriously
: consider it for 2.6.37 inclusion even at this late date).
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reported-by: Shaohua Li <shaohua.li@intel.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Tested-by: Nicolas Bareil <nico@chdir.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Kyle McMartin <kyle@mcmartin.ca>
Cc: <stable@kernel.org> [2.6.37.1, 2.6.36.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-13 23:45:41 +00:00
|
|
|
|
2020-06-03 22:59:01 +00:00
|
|
|
trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
|
2018-04-05 23:25:16 +00:00
|
|
|
gfp_flags);
|
2005-09-13 08:25:07 +00:00
|
|
|
wake_up_interruptible(&pgdat->kswapd_wait);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2009-05-24 20:16:31 +00:00
|
|
|
#ifdef CONFIG_HIBERNATION
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
vmscan: kill hibernation specific reclaim logic and unify it
shrink_all_zone() was introduced by commit d6277db4ab (swsusp: rework
memory shrinker) for hibernate performance improvement. and
sc.swap_cluster_max was introduced by commit a06fe4d307 (Speed freeing
memory for suspend).
commit a06fe4d307 said
Without the patch:
Freed 14600 pages in 1749 jiffies = 32.61 MB/s (Anomolous!)
Freed 88563 pages in 14719 jiffies = 23.50 MB/s
Freed 205734 pages in 32389 jiffies = 24.81 MB/s
With the patch:
Freed 68252 pages in 496 jiffies = 537.52 MB/s
Freed 116464 pages in 569 jiffies = 798.54 MB/s
Freed 209699 pages in 705 jiffies = 1161.89 MB/s
At that time, their patch was pretty worth. However, Modern Hardware
trend and recent VM improvement broke its worth. From several reason, I
think we should remove shrink_all_zones() at all.
detail:
1) Old days, shrink_zone()'s slowness was mainly caused by stupid io-throttle
at no i/o congestion.
but current shrink_zone() is sane, not slow.
2) shrink_all_zone() try to shrink all pages at a time. but it doesn't works
fine on numa system.
example)
System has 4GB memory and each node have 2GB. and hibernate need 1GB.
optimal)
steal 500MB from each node.
shrink_all_zones)
steal 1GB from node-0.
Oh, Cache balancing logic was broken. ;)
Unfortunately, Desktop system moved ahead NUMA at nowadays.
(Side note, if hibernate require 2GB, shrink_all_zones() never success
on above machine)
3) if the node has several I/O flighting pages, shrink_all_zones() makes
pretty bad result.
schenario) hibernate need 1GB
1) shrink_all_zones() try to reclaim 1GB from Node-0
2) but it only reclaimed 990MB
3) stupidly, shrink_all_zones() try to reclaim 1GB from Node-1
4) it reclaimed 990MB
Oh, well. it reclaimed twice much than required.
In the other hand, current shrink_zone() has sane baling out logic.
then, it doesn't make overkill reclaim. then, we lost shrink_zones()'s risk.
4) SplitLRU VM always keep active/inactive ratio very carefully. inactive list only
shrinking break its assumption. it makes unnecessary OOM risk. it obviously suboptimal.
Now, shrink_all_memory() is only the wrapper function of do_try_to_free_pages().
it bring good reviewability and debuggability, and solve above problems.
side note: Reclaim logic unificication makes two good side effect.
- Fix recursive reclaim bug on shrink_all_memory().
it did forgot to use PF_MEMALLOC. it mean the system be able to stuck into deadlock.
- Now, shrink_all_memory() got lockdep awareness. it bring good debuggability.
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:59:12 +00:00
|
|
|
* Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
|
2006-06-23 09:03:18 +00:00
|
|
|
* freed pages.
|
|
|
|
*
|
|
|
|
* Rather than trying to age LRUs the aim is to preserve the overall
|
|
|
|
* LRU order by reclaiming preferentially
|
|
|
|
* inactive > active > active referenced > active mapped
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
vmscan: kill hibernation specific reclaim logic and unify it
shrink_all_zone() was introduced by commit d6277db4ab (swsusp: rework
memory shrinker) for hibernate performance improvement. and
sc.swap_cluster_max was introduced by commit a06fe4d307 (Speed freeing
memory for suspend).
commit a06fe4d307 said
Without the patch:
Freed 14600 pages in 1749 jiffies = 32.61 MB/s (Anomolous!)
Freed 88563 pages in 14719 jiffies = 23.50 MB/s
Freed 205734 pages in 32389 jiffies = 24.81 MB/s
With the patch:
Freed 68252 pages in 496 jiffies = 537.52 MB/s
Freed 116464 pages in 569 jiffies = 798.54 MB/s
Freed 209699 pages in 705 jiffies = 1161.89 MB/s
At that time, their patch was pretty worth. However, Modern Hardware
trend and recent VM improvement broke its worth. From several reason, I
think we should remove shrink_all_zones() at all.
detail:
1) Old days, shrink_zone()'s slowness was mainly caused by stupid io-throttle
at no i/o congestion.
but current shrink_zone() is sane, not slow.
2) shrink_all_zone() try to shrink all pages at a time. but it doesn't works
fine on numa system.
example)
System has 4GB memory and each node have 2GB. and hibernate need 1GB.
optimal)
steal 500MB from each node.
shrink_all_zones)
steal 1GB from node-0.
Oh, Cache balancing logic was broken. ;)
Unfortunately, Desktop system moved ahead NUMA at nowadays.
(Side note, if hibernate require 2GB, shrink_all_zones() never success
on above machine)
3) if the node has several I/O flighting pages, shrink_all_zones() makes
pretty bad result.
schenario) hibernate need 1GB
1) shrink_all_zones() try to reclaim 1GB from Node-0
2) but it only reclaimed 990MB
3) stupidly, shrink_all_zones() try to reclaim 1GB from Node-1
4) it reclaimed 990MB
Oh, well. it reclaimed twice much than required.
In the other hand, current shrink_zone() has sane baling out logic.
then, it doesn't make overkill reclaim. then, we lost shrink_zones()'s risk.
4) SplitLRU VM always keep active/inactive ratio very carefully. inactive list only
shrinking break its assumption. it makes unnecessary OOM risk. it obviously suboptimal.
Now, shrink_all_memory() is only the wrapper function of do_try_to_free_pages().
it bring good reviewability and debuggability, and solve above problems.
side note: Reclaim logic unificication makes two good side effect.
- Fix recursive reclaim bug on shrink_all_memory().
it did forgot to use PF_MEMALLOC. it mean the system be able to stuck into deadlock.
- Now, shrink_all_memory() got lockdep awareness. it bring good debuggability.
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:59:12 +00:00
|
|
|
unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2006-06-23 09:03:18 +00:00
|
|
|
struct scan_control sc = {
|
2014-08-06 23:06:19 +00:00
|
|
|
.nr_to_reclaim = nr_to_reclaim,
|
vmscan: kill hibernation specific reclaim logic and unify it
shrink_all_zone() was introduced by commit d6277db4ab (swsusp: rework
memory shrinker) for hibernate performance improvement. and
sc.swap_cluster_max was introduced by commit a06fe4d307 (Speed freeing
memory for suspend).
commit a06fe4d307 said
Without the patch:
Freed 14600 pages in 1749 jiffies = 32.61 MB/s (Anomolous!)
Freed 88563 pages in 14719 jiffies = 23.50 MB/s
Freed 205734 pages in 32389 jiffies = 24.81 MB/s
With the patch:
Freed 68252 pages in 496 jiffies = 537.52 MB/s
Freed 116464 pages in 569 jiffies = 798.54 MB/s
Freed 209699 pages in 705 jiffies = 1161.89 MB/s
At that time, their patch was pretty worth. However, Modern Hardware
trend and recent VM improvement broke its worth. From several reason, I
think we should remove shrink_all_zones() at all.
detail:
1) Old days, shrink_zone()'s slowness was mainly caused by stupid io-throttle
at no i/o congestion.
but current shrink_zone() is sane, not slow.
2) shrink_all_zone() try to shrink all pages at a time. but it doesn't works
fine on numa system.
example)
System has 4GB memory and each node have 2GB. and hibernate need 1GB.
optimal)
steal 500MB from each node.
shrink_all_zones)
steal 1GB from node-0.
Oh, Cache balancing logic was broken. ;)
Unfortunately, Desktop system moved ahead NUMA at nowadays.
(Side note, if hibernate require 2GB, shrink_all_zones() never success
on above machine)
3) if the node has several I/O flighting pages, shrink_all_zones() makes
pretty bad result.
schenario) hibernate need 1GB
1) shrink_all_zones() try to reclaim 1GB from Node-0
2) but it only reclaimed 990MB
3) stupidly, shrink_all_zones() try to reclaim 1GB from Node-1
4) it reclaimed 990MB
Oh, well. it reclaimed twice much than required.
In the other hand, current shrink_zone() has sane baling out logic.
then, it doesn't make overkill reclaim. then, we lost shrink_zones()'s risk.
4) SplitLRU VM always keep active/inactive ratio very carefully. inactive list only
shrinking break its assumption. it makes unnecessary OOM risk. it obviously suboptimal.
Now, shrink_all_memory() is only the wrapper function of do_try_to_free_pages().
it bring good reviewability and debuggability, and solve above problems.
side note: Reclaim logic unificication makes two good side effect.
- Fix recursive reclaim bug on shrink_all_memory().
it did forgot to use PF_MEMALLOC. it mean the system be able to stuck into deadlock.
- Now, shrink_all_memory() got lockdep awareness. it bring good debuggability.
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:59:12 +00:00
|
|
|
.gfp_mask = GFP_HIGHUSER_MOVABLE,
|
2016-07-28 22:45:37 +00:00
|
|
|
.reclaim_idx = MAX_NR_ZONES - 1,
|
2014-08-06 23:06:19 +00:00
|
|
|
.priority = DEF_PRIORITY,
|
2006-06-23 09:03:18 +00:00
|
|
|
.may_writepage = 1,
|
2014-08-06 23:06:19 +00:00
|
|
|
.may_unmap = 1,
|
|
|
|
.may_swap = 1,
|
vmscan: kill hibernation specific reclaim logic and unify it
shrink_all_zone() was introduced by commit d6277db4ab (swsusp: rework
memory shrinker) for hibernate performance improvement. and
sc.swap_cluster_max was introduced by commit a06fe4d307 (Speed freeing
memory for suspend).
commit a06fe4d307 said
Without the patch:
Freed 14600 pages in 1749 jiffies = 32.61 MB/s (Anomolous!)
Freed 88563 pages in 14719 jiffies = 23.50 MB/s
Freed 205734 pages in 32389 jiffies = 24.81 MB/s
With the patch:
Freed 68252 pages in 496 jiffies = 537.52 MB/s
Freed 116464 pages in 569 jiffies = 798.54 MB/s
Freed 209699 pages in 705 jiffies = 1161.89 MB/s
At that time, their patch was pretty worth. However, Modern Hardware
trend and recent VM improvement broke its worth. From several reason, I
think we should remove shrink_all_zones() at all.
detail:
1) Old days, shrink_zone()'s slowness was mainly caused by stupid io-throttle
at no i/o congestion.
but current shrink_zone() is sane, not slow.
2) shrink_all_zone() try to shrink all pages at a time. but it doesn't works
fine on numa system.
example)
System has 4GB memory and each node have 2GB. and hibernate need 1GB.
optimal)
steal 500MB from each node.
shrink_all_zones)
steal 1GB from node-0.
Oh, Cache balancing logic was broken. ;)
Unfortunately, Desktop system moved ahead NUMA at nowadays.
(Side note, if hibernate require 2GB, shrink_all_zones() never success
on above machine)
3) if the node has several I/O flighting pages, shrink_all_zones() makes
pretty bad result.
schenario) hibernate need 1GB
1) shrink_all_zones() try to reclaim 1GB from Node-0
2) but it only reclaimed 990MB
3) stupidly, shrink_all_zones() try to reclaim 1GB from Node-1
4) it reclaimed 990MB
Oh, well. it reclaimed twice much than required.
In the other hand, current shrink_zone() has sane baling out logic.
then, it doesn't make overkill reclaim. then, we lost shrink_zones()'s risk.
4) SplitLRU VM always keep active/inactive ratio very carefully. inactive list only
shrinking break its assumption. it makes unnecessary OOM risk. it obviously suboptimal.
Now, shrink_all_memory() is only the wrapper function of do_try_to_free_pages().
it bring good reviewability and debuggability, and solve above problems.
side note: Reclaim logic unificication makes two good side effect.
- Fix recursive reclaim bug on shrink_all_memory().
it did forgot to use PF_MEMALLOC. it mean the system be able to stuck into deadlock.
- Now, shrink_all_memory() got lockdep awareness. it bring good debuggability.
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:59:12 +00:00
|
|
|
.hibernation_mode = 1,
|
2005-04-16 22:20:36 +00:00
|
|
|
};
|
2011-05-25 00:12:26 +00:00
|
|
|
struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
|
vmscan: kill hibernation specific reclaim logic and unify it
shrink_all_zone() was introduced by commit d6277db4ab (swsusp: rework
memory shrinker) for hibernate performance improvement. and
sc.swap_cluster_max was introduced by commit a06fe4d307 (Speed freeing
memory for suspend).
commit a06fe4d307 said
Without the patch:
Freed 14600 pages in 1749 jiffies = 32.61 MB/s (Anomolous!)
Freed 88563 pages in 14719 jiffies = 23.50 MB/s
Freed 205734 pages in 32389 jiffies = 24.81 MB/s
With the patch:
Freed 68252 pages in 496 jiffies = 537.52 MB/s
Freed 116464 pages in 569 jiffies = 798.54 MB/s
Freed 209699 pages in 705 jiffies = 1161.89 MB/s
At that time, their patch was pretty worth. However, Modern Hardware
trend and recent VM improvement broke its worth. From several reason, I
think we should remove shrink_all_zones() at all.
detail:
1) Old days, shrink_zone()'s slowness was mainly caused by stupid io-throttle
at no i/o congestion.
but current shrink_zone() is sane, not slow.
2) shrink_all_zone() try to shrink all pages at a time. but it doesn't works
fine on numa system.
example)
System has 4GB memory and each node have 2GB. and hibernate need 1GB.
optimal)
steal 500MB from each node.
shrink_all_zones)
steal 1GB from node-0.
Oh, Cache balancing logic was broken. ;)
Unfortunately, Desktop system moved ahead NUMA at nowadays.
(Side note, if hibernate require 2GB, shrink_all_zones() never success
on above machine)
3) if the node has several I/O flighting pages, shrink_all_zones() makes
pretty bad result.
schenario) hibernate need 1GB
1) shrink_all_zones() try to reclaim 1GB from Node-0
2) but it only reclaimed 990MB
3) stupidly, shrink_all_zones() try to reclaim 1GB from Node-1
4) it reclaimed 990MB
Oh, well. it reclaimed twice much than required.
In the other hand, current shrink_zone() has sane baling out logic.
then, it doesn't make overkill reclaim. then, we lost shrink_zones()'s risk.
4) SplitLRU VM always keep active/inactive ratio very carefully. inactive list only
shrinking break its assumption. it makes unnecessary OOM risk. it obviously suboptimal.
Now, shrink_all_memory() is only the wrapper function of do_try_to_free_pages().
it bring good reviewability and debuggability, and solve above problems.
side note: Reclaim logic unificication makes two good side effect.
- Fix recursive reclaim bug on shrink_all_memory().
it did forgot to use PF_MEMALLOC. it mean the system be able to stuck into deadlock.
- Now, shrink_all_memory() got lockdep awareness. it bring good debuggability.
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:59:12 +00:00
|
|
|
unsigned long nr_reclaimed;
|
2017-05-08 22:59:50 +00:00
|
|
|
unsigned int noreclaim_flag;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2017-03-03 09:13:38 +00:00
|
|
|
fs_reclaim_acquire(sc.gfp_mask);
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
noreclaim_flag = memalloc_noreclaim_save();
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(current, &sc.reclaim_state);
|
2006-06-23 09:03:18 +00:00
|
|
|
|
2014-04-03 21:47:22 +00:00
|
|
|
nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
|
2009-03-31 22:19:34 +00:00
|
|
|
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(current, NULL);
|
2017-05-08 22:59:50 +00:00
|
|
|
memalloc_noreclaim_restore(noreclaim_flag);
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
fs_reclaim_release(sc.gfp_mask);
|
2006-06-23 09:03:18 +00:00
|
|
|
|
vmscan: kill hibernation specific reclaim logic and unify it
shrink_all_zone() was introduced by commit d6277db4ab (swsusp: rework
memory shrinker) for hibernate performance improvement. and
sc.swap_cluster_max was introduced by commit a06fe4d307 (Speed freeing
memory for suspend).
commit a06fe4d307 said
Without the patch:
Freed 14600 pages in 1749 jiffies = 32.61 MB/s (Anomolous!)
Freed 88563 pages in 14719 jiffies = 23.50 MB/s
Freed 205734 pages in 32389 jiffies = 24.81 MB/s
With the patch:
Freed 68252 pages in 496 jiffies = 537.52 MB/s
Freed 116464 pages in 569 jiffies = 798.54 MB/s
Freed 209699 pages in 705 jiffies = 1161.89 MB/s
At that time, their patch was pretty worth. However, Modern Hardware
trend and recent VM improvement broke its worth. From several reason, I
think we should remove shrink_all_zones() at all.
detail:
1) Old days, shrink_zone()'s slowness was mainly caused by stupid io-throttle
at no i/o congestion.
but current shrink_zone() is sane, not slow.
2) shrink_all_zone() try to shrink all pages at a time. but it doesn't works
fine on numa system.
example)
System has 4GB memory and each node have 2GB. and hibernate need 1GB.
optimal)
steal 500MB from each node.
shrink_all_zones)
steal 1GB from node-0.
Oh, Cache balancing logic was broken. ;)
Unfortunately, Desktop system moved ahead NUMA at nowadays.
(Side note, if hibernate require 2GB, shrink_all_zones() never success
on above machine)
3) if the node has several I/O flighting pages, shrink_all_zones() makes
pretty bad result.
schenario) hibernate need 1GB
1) shrink_all_zones() try to reclaim 1GB from Node-0
2) but it only reclaimed 990MB
3) stupidly, shrink_all_zones() try to reclaim 1GB from Node-1
4) it reclaimed 990MB
Oh, well. it reclaimed twice much than required.
In the other hand, current shrink_zone() has sane baling out logic.
then, it doesn't make overkill reclaim. then, we lost shrink_zones()'s risk.
4) SplitLRU VM always keep active/inactive ratio very carefully. inactive list only
shrinking break its assumption. it makes unnecessary OOM risk. it obviously suboptimal.
Now, shrink_all_memory() is only the wrapper function of do_try_to_free_pages().
it bring good reviewability and debuggability, and solve above problems.
side note: Reclaim logic unificication makes two good side effect.
- Fix recursive reclaim bug on shrink_all_memory().
it did forgot to use PF_MEMALLOC. it mean the system be able to stuck into deadlock.
- Now, shrink_all_memory() got lockdep awareness. it bring good debuggability.
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:59:12 +00:00
|
|
|
return nr_reclaimed;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2009-05-24 20:16:31 +00:00
|
|
|
#endif /* CONFIG_HIBERNATION */
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2006-06-27 09:53:33 +00:00
|
|
|
/*
|
|
|
|
* This kswapd start function will be called by init and node-hot-add.
|
|
|
|
*/
|
2021-09-02 21:59:46 +00:00
|
|
|
void kswapd_run(int nid)
|
2006-06-27 09:53:33 +00:00
|
|
|
{
|
|
|
|
pg_data_t *pgdat = NODE_DATA(nid);
|
|
|
|
|
2022-08-27 11:19:59 +00:00
|
|
|
pgdat_kswapd_lock(pgdat);
|
|
|
|
if (!pgdat->kswapd) {
|
|
|
|
pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
|
|
|
|
if (IS_ERR(pgdat->kswapd)) {
|
|
|
|
/* failure at boot is fatal */
|
|
|
|
BUG_ON(system_state < SYSTEM_RUNNING);
|
|
|
|
pr_err("Failed to start kswapd on node %d\n", nid);
|
|
|
|
pgdat->kswapd = NULL;
|
|
|
|
}
|
2006-06-27 09:53:33 +00:00
|
|
|
}
|
2022-08-27 11:19:59 +00:00
|
|
|
pgdat_kswapd_unlock(pgdat);
|
2006-06-27 09:53:33 +00:00
|
|
|
}
|
|
|
|
|
2009-12-15 01:58:33 +00:00
|
|
|
/*
|
2012-07-11 21:01:52 +00:00
|
|
|
* Called by memory hotplug when all memory in a node is offlined. Caller must
|
2022-06-20 07:15:16 +00:00
|
|
|
* be holding mem_hotplug_begin/done().
|
2009-12-15 01:58:33 +00:00
|
|
|
*/
|
|
|
|
void kswapd_stop(int nid)
|
|
|
|
{
|
2022-08-27 11:19:59 +00:00
|
|
|
pg_data_t *pgdat = NODE_DATA(nid);
|
|
|
|
struct task_struct *kswapd;
|
2009-12-15 01:58:33 +00:00
|
|
|
|
2022-08-27 11:19:59 +00:00
|
|
|
pgdat_kswapd_lock(pgdat);
|
|
|
|
kswapd = pgdat->kswapd;
|
2012-07-11 21:01:52 +00:00
|
|
|
if (kswapd) {
|
2009-12-15 01:58:33 +00:00
|
|
|
kthread_stop(kswapd);
|
2022-08-27 11:19:59 +00:00
|
|
|
pgdat->kswapd = NULL;
|
2012-07-11 21:01:52 +00:00
|
|
|
}
|
2022-08-27 11:19:59 +00:00
|
|
|
pgdat_kswapd_unlock(pgdat);
|
2009-12-15 01:58:33 +00:00
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
static int __init kswapd_init(void)
|
|
|
|
{
|
2020-04-02 04:10:09 +00:00
|
|
|
int nid;
|
2006-03-22 08:08:19 +00:00
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
swap_setup();
|
2012-12-12 21:51:43 +00:00
|
|
|
for_each_node_state(nid, N_MEMORY)
|
2006-06-27 09:53:33 +00:00
|
|
|
kswapd_run(nid);
|
2005-04-16 22:20:36 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
module_init(kswapd_init)
|
2006-01-19 01:42:31 +00:00
|
|
|
|
|
|
|
#ifdef CONFIG_NUMA
|
|
|
|
/*
|
2016-07-28 22:46:32 +00:00
|
|
|
* Node reclaim mode
|
2006-01-19 01:42:31 +00:00
|
|
|
*
|
2016-07-28 22:46:32 +00:00
|
|
|
* If non-zero call node_reclaim when the number of free pages falls below
|
2006-01-19 01:42:31 +00:00
|
|
|
* the watermarks.
|
|
|
|
*/
|
2016-07-28 22:46:32 +00:00
|
|
|
int node_reclaim_mode __read_mostly;
|
2006-01-19 01:42:31 +00:00
|
|
|
|
2006-02-01 11:05:32 +00:00
|
|
|
/*
|
2016-07-28 22:46:32 +00:00
|
|
|
* Priority for NODE_RECLAIM. This determines the fraction of pages
|
2006-02-01 11:05:32 +00:00
|
|
|
* of a node considered for each zone_reclaim. 4 scans 1/16th of
|
|
|
|
* a zone.
|
|
|
|
*/
|
2016-07-28 22:46:32 +00:00
|
|
|
#define NODE_RECLAIM_PRIORITY 4
|
2006-02-01 11:05:32 +00:00
|
|
|
|
2006-07-03 07:24:13 +00:00
|
|
|
/*
|
2016-07-28 22:46:32 +00:00
|
|
|
* Percentage of pages in a zone that must be unmapped for node_reclaim to
|
2006-07-03 07:24:13 +00:00
|
|
|
* occur.
|
|
|
|
*/
|
|
|
|
int sysctl_min_unmapped_ratio = 1;
|
|
|
|
|
2006-09-26 06:31:52 +00:00
|
|
|
/*
|
|
|
|
* If the number of slab pages in a zone grows beyond this percentage then
|
|
|
|
* slab reclaim needs to occur.
|
|
|
|
*/
|
|
|
|
int sysctl_min_slab_ratio = 5;
|
|
|
|
|
2016-07-28 22:46:20 +00:00
|
|
|
static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
{
|
2016-07-28 22:46:20 +00:00
|
|
|
unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
|
|
|
|
unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
|
|
|
|
node_page_state(pgdat, NR_ACTIVE_FILE);
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* It's possible for there to be more file mapped pages than
|
|
|
|
* accounted for by the pages on the file LRU lists because
|
|
|
|
* tmpfs pages accounted for as ANON can also be FILE_MAPPED
|
|
|
|
*/
|
|
|
|
return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Work out how many page cache pages we can reclaim in this reclaim_mode */
|
2016-07-28 22:46:32 +00:00
|
|
|
static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
{
|
2015-11-06 02:48:08 +00:00
|
|
|
unsigned long nr_pagecache_reclaimable;
|
|
|
|
unsigned long delta = 0;
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
|
|
|
|
/*
|
2015-06-24 23:56:42 +00:00
|
|
|
* If RECLAIM_UNMAP is set, then all file pages are considered
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
* potentially reclaimable. Otherwise, we have to worry about
|
2016-07-28 22:46:20 +00:00
|
|
|
* pages like swapcache and node_unmapped_file_pages() provides
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
* a better estimate
|
|
|
|
*/
|
2016-07-28 22:46:32 +00:00
|
|
|
if (node_reclaim_mode & RECLAIM_UNMAP)
|
|
|
|
nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
else
|
2016-07-28 22:46:32 +00:00
|
|
|
nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
|
|
|
|
/* If we can't clean pages, remove dirty pages from consideration */
|
2016-07-28 22:46:32 +00:00
|
|
|
if (!(node_reclaim_mode & RECLAIM_WRITE))
|
|
|
|
delta += node_page_state(pgdat, NR_FILE_DIRTY);
|
vmscan: properly account for the number of page cache pages zone_reclaim() can reclaim
A bug was brought to my attention against a distro kernel but it affects
mainline and I believe problems like this have been reported in various
guises on the mailing lists although I don't have specific examples at the
moment.
The reported problem was that malloc() stalled for a long time (minutes in
some cases) if a large tmpfs mount was occupying a large percentage of
memory overall. The pages did not get cleaned or reclaimed by
zone_reclaim() because the zone_reclaim_mode was unsuitable, but the lists
are uselessly scanned frequencly making the CPU spin at near 100%.
This patchset intends to address that bug and bring the behaviour of
zone_reclaim() more in line with expectations which were noticed during
investigation. It is based on top of mmotm and takes advantage of
Kosaki's work with respect to zone_reclaim().
Patch 1 fixes the heuristics that zone_reclaim() uses to determine if the
scan should go ahead. The broken heuristic is what was causing the
malloc() stall as it uselessly scanned the LRU constantly. Currently,
zone_reclaim is assuming zone_reclaim_mode is 1 and historically it
could not deal with tmpfs pages at all. This fixes up the heuristic so
that an unnecessary scan is more likely to be correctly avoided.
Patch 2 notes that zone_reclaim() returning a failure automatically means
the zone is marked full. This is not always true. It could have
failed because the GFP mask or zone_reclaim_mode were unsuitable.
Patch 3 introduces a counter zreclaim_failed that will increment each
time the zone_reclaim scan-avoidance heuristics fail. If that
counter is rapidly increasing, then zone_reclaim_mode should be
set to 0 as a temporarily resolution and a bug reported because
the scan-avoidance heuristic is still broken.
This patch:
On NUMA machines, the administrator can configure zone_reclaim_mode that
is a more targetted form of direct reclaim. On machines with large NUMA
distances for example, a zone_reclaim_mode defaults to 1 meaning that
clean unmapped pages will be reclaimed if the zone watermarks are not
being met.
There is a heuristic that determines if the scan is worthwhile but the
problem is that the heuristic is not being properly applied and is
basically assuming zone_reclaim_mode is 1 if it is enabled. The lack of
proper detection can manfiest as high CPU usage as the LRU list is scanned
uselessly.
Historically, once enabled it was depending on NR_FILE_PAGES which may
include swapcache pages that the reclaim_mode cannot deal with. Patch
vmscan-change-the-number-of-the-unmapped-files-in-zone-reclaim.patch by
Kosaki Motohiro noted that zone_page_state(zone, NR_FILE_PAGES) included
pages that were not file-backed such as swapcache and made a calculation
based on the inactive, active and mapped files. This is far superior when
zone_reclaim==1 but if RECLAIM_SWAP is set, then NR_FILE_PAGES is a
reasonable starting figure.
This patch alters how zone_reclaim() works out how many pages it might be
able to reclaim given the current reclaim_mode. If RECLAIM_SWAP is set in
the reclaim_mode it will either consider NR_FILE_PAGES as potential
candidates or else use NR_{IN}ACTIVE}_PAGES-NR_FILE_MAPPED to discount
swapcache and other non-file-backed pages. If RECLAIM_WRITE is not set,
then NR_FILE_DIRTY number of pages are not candidates. If RECLAIM_SWAP is
not set, then NR_FILE_MAPPED are not.
[kosaki.motohiro@jp.fujitsu.com: Estimate unmapped pages minus tmpfs pages]
[fengguang.wu@intel.com: Fix underflow problem in Kosaki's estimate]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Reviewed-by: Rik van Riel <riel@redhat.com>
Acked-by: Christoph Lameter <cl@linux-foundation.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-06-16 22:33:20 +00:00
|
|
|
|
|
|
|
/* Watch for any possible underflows due to delta */
|
|
|
|
if (unlikely(delta > nr_pagecache_reclaimable))
|
|
|
|
delta = nr_pagecache_reclaimable;
|
|
|
|
|
|
|
|
return nr_pagecache_reclaimable - delta;
|
|
|
|
}
|
|
|
|
|
2006-01-19 01:42:31 +00:00
|
|
|
/*
|
2016-07-28 22:46:32 +00:00
|
|
|
* Try to free up some pages from this node through reclaim.
|
2006-01-19 01:42:31 +00:00
|
|
|
*/
|
2016-07-28 22:46:32 +00:00
|
|
|
static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
|
2006-01-19 01:42:31 +00:00
|
|
|
{
|
2006-03-22 08:08:22 +00:00
|
|
|
/* Minimum pages needed in order to stay on node */
|
2006-03-22 08:08:19 +00:00
|
|
|
const unsigned long nr_pages = 1 << order;
|
2006-01-19 01:42:31 +00:00
|
|
|
struct task_struct *p = current;
|
2017-05-08 22:59:50 +00:00
|
|
|
unsigned int noreclaim_flag;
|
2006-03-22 08:08:18 +00:00
|
|
|
struct scan_control sc = {
|
2013-02-23 00:32:24 +00:00
|
|
|
.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
|
2017-07-06 22:36:50 +00:00
|
|
|
.gfp_mask = current_gfp_context(gfp_mask),
|
2009-03-31 22:19:38 +00:00
|
|
|
.order = order,
|
2016-07-28 22:46:32 +00:00
|
|
|
.priority = NODE_RECLAIM_PRIORITY,
|
|
|
|
.may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
|
|
|
|
.may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
|
2014-08-06 23:06:19 +00:00
|
|
|
.may_swap = 1,
|
2017-07-06 22:36:50 +00:00
|
|
|
.reclaim_idx = gfp_zone(gfp_mask),
|
2006-03-22 08:08:18 +00:00
|
|
|
};
|
2021-08-20 02:04:27 +00:00
|
|
|
unsigned long pflags;
|
2006-01-19 01:42:31 +00:00
|
|
|
|
2019-05-14 00:17:53 +00:00
|
|
|
trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
|
|
|
|
sc.gfp_mask);
|
|
|
|
|
2006-01-19 01:42:31 +00:00
|
|
|
cond_resched();
|
2021-08-20 02:04:27 +00:00
|
|
|
psi_memstall_enter(&pflags);
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
fs_reclaim_acquire(sc.gfp_mask);
|
2006-02-24 21:04:22 +00:00
|
|
|
/*
|
2015-06-24 23:56:42 +00:00
|
|
|
* We need to be able to allocate from the reserves for RECLAIM_UNMAP
|
2006-02-24 21:04:22 +00:00
|
|
|
*/
|
2017-05-08 22:59:50 +00:00
|
|
|
noreclaim_flag = memalloc_noreclaim_save();
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(p, &sc.reclaim_state);
|
2006-02-01 11:05:29 +00:00
|
|
|
|
2022-05-13 03:22:58 +00:00
|
|
|
if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
|
|
|
|
node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
|
2006-09-26 06:31:52 +00:00
|
|
|
/*
|
2018-04-10 23:27:51 +00:00
|
|
|
* Free memory by calling shrink node with increasing
|
2006-09-26 06:31:52 +00:00
|
|
|
* priorities until we have enough memory freed.
|
|
|
|
*/
|
|
|
|
do {
|
2016-07-28 22:46:35 +00:00
|
|
|
shrink_node(pgdat, &sc);
|
2012-05-29 22:06:57 +00:00
|
|
|
} while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
|
2006-09-26 06:31:52 +00:00
|
|
|
}
|
2006-02-01 11:05:29 +00:00
|
|
|
|
2019-07-16 23:26:15 +00:00
|
|
|
set_task_reclaim_state(p, NULL);
|
2017-05-08 22:59:50 +00:00
|
|
|
memalloc_noreclaim_restore(noreclaim_flag);
|
lockdep: fix fs_reclaim annotation
While revisiting my Btrfs swapfile series [1], I introduced a situation
in which reclaim would lock i_rwsem, and even though the swapon() path
clearly made GFP_KERNEL allocations while holding i_rwsem, I got no
complaints from lockdep. It turns out that the rework of the fs_reclaim
annotation was broken: if the current task has PF_MEMALLOC set, we don't
acquire the dummy fs_reclaim lock, but when reclaiming we always check
this _after_ we've just set the PF_MEMALLOC flag. In most cases, we can
fix this by moving the fs_reclaim_{acquire,release}() outside of the
memalloc_noreclaim_{save,restore}(), althought kswapd is slightly
different. After applying this, I got the expected lockdep splats.
1: https://lwn.net/Articles/625412/
Link: http://lkml.kernel.org/r/9f8aa70652a98e98d7c4de0fc96a4addcee13efe.1523778026.git.osandov@fb.com
Fixes: d92a8cfcb37e ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-08 00:07:02 +00:00
|
|
|
fs_reclaim_release(sc.gfp_mask);
|
2021-08-20 02:04:27 +00:00
|
|
|
psi_memstall_leave(&pflags);
|
2019-05-14 00:17:53 +00:00
|
|
|
|
|
|
|
trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
|
|
|
|
|
vmscan: bail out of direct reclaim after swap_cluster_max pages
When the VM is under pressure, it can happen that several direct reclaim
processes are in the pageout code simultaneously. It also happens that
the reclaiming processes run into mostly referenced, mapped and dirty
pages in the first round.
This results in multiple direct reclaim processes having a lower
pageout priority, which corresponds to a higher target of pages to
scan.
This in turn can result in each direct reclaim process freeing
many pages. Together, they can end up freeing way too many pages.
This kicks useful data out of memory (in some cases more than half
of all memory is swapped out). It also impacts performance by
keeping tasks stuck in the pageout code for too long.
A 30% improvement in hackbench has been observed with this patch.
The fix is relatively simple: in shrink_zone() we can check how many
pages we have already freed, direct reclaim tasks break out of the
scanning loop if they have already freed enough pages and have reached
a lower priority level.
We do not break out of shrink_zone() when priority == DEF_PRIORITY,
to ensure that equal pressure is applied to every zone in the common
case.
However, in order to do this we do need to know how many pages we already
freed, so move nr_reclaimed into scan_control.
akpm: a historical interlude...
We tried this in 2004:
:commit e468e46a9bea3297011d5918663ce6d19094cf87
:Author: akpm <akpm>
:Date: Thu Jun 24 15:53:52 2004 +0000
:
:[PATCH] vmscan.c: dont reclaim too many pages
:
: The shrink_zone() logic can, under some circumstances, cause far too many
: pages to be reclaimed. Say, we're scanning at high priority and suddenly hit
: a large number of reclaimable pages on the LRU.
: Change things so we bale out when SWAP_CLUSTER_MAX pages have been reclaimed.
And we reverted it in 2006:
:commit 210fe530305ee50cd889fe9250168228b2994f32
:Author: Andrew Morton <akpm@osdl.org>
:Date: Fri Jan 6 00:11:14 2006 -0800
:
: [PATCH] vmscan: balancing fix
:
: Revert a patch which went into 2.6.8-rc1. The changelog for that patch was:
:
: The shrink_zone() logic can, under some circumstances, cause far too many
: pages to be reclaimed. Say, we're scanning at high priority and suddenly
: hit a large number of reclaimable pages on the LRU.
:
: Change things so we bale out when SWAP_CLUSTER_MAX pages have been
: reclaimed.
:
: Problem is, this change caused significant imbalance in inter-zone scan
: balancing by truncating scans of larger zones.
:
: Suppose, for example, ZONE_HIGHMEM is 10x the size of ZONE_NORMAL. The zone
: balancing algorithm would require that if we're scanning 100 pages of
: ZONE_HIGHMEM, we should scan 10 pages of ZONE_NORMAL. But this logic will
: cause the scanning of ZONE_HIGHMEM to bale out after only 32 pages are
: reclaimed. Thus effectively causing smaller zones to be scanned relatively
: harder than large ones.
:
: Now I need to remember what the workload was which caused me to write this
: patch originally, then fix it up in a different way...
And we haven't demonstrated that whatever problem caused that reversion is
not being reintroduced by this change in 2008.
Signed-off-by: Rik van Riel <riel@redhat.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:40:01 +00:00
|
|
|
return sc.nr_reclaimed >= nr_pages;
|
2006-01-19 01:42:31 +00:00
|
|
|
}
|
2006-03-22 08:08:18 +00:00
|
|
|
|
2016-07-28 22:46:32 +00:00
|
|
|
int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
|
2006-03-22 08:08:18 +00:00
|
|
|
{
|
2007-10-17 06:26:01 +00:00
|
|
|
int ret;
|
2006-03-22 08:08:18 +00:00
|
|
|
|
|
|
|
/*
|
2016-07-28 22:46:32 +00:00
|
|
|
* Node reclaim reclaims unmapped file backed pages and
|
2006-09-26 06:31:52 +00:00
|
|
|
* slab pages if we are over the defined limits.
|
2006-06-30 08:55:37 +00:00
|
|
|
*
|
2006-07-03 07:24:13 +00:00
|
|
|
* A small portion of unmapped file backed pages is needed for
|
|
|
|
* file I/O otherwise pages read by file I/O will be immediately
|
2016-07-28 22:46:32 +00:00
|
|
|
* thrown out if the node is overallocated. So we do not reclaim
|
|
|
|
* if less than a specified percentage of the node is used by
|
2006-07-03 07:24:13 +00:00
|
|
|
* unmapped file backed pages.
|
2006-03-22 08:08:18 +00:00
|
|
|
*/
|
2016-07-28 22:46:32 +00:00
|
|
|
if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
|
2020-08-07 06:20:39 +00:00
|
|
|
node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
|
|
|
|
pgdat->min_slab_pages)
|
2016-07-28 22:46:32 +00:00
|
|
|
return NODE_RECLAIM_FULL;
|
2006-03-22 08:08:18 +00:00
|
|
|
|
|
|
|
/*
|
2007-10-17 06:26:01 +00:00
|
|
|
* Do not scan if the allocation should not be delayed.
|
2006-03-22 08:08:18 +00:00
|
|
|
*/
|
2015-11-07 00:28:21 +00:00
|
|
|
if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
|
2016-07-28 22:46:32 +00:00
|
|
|
return NODE_RECLAIM_NOSCAN;
|
2006-03-22 08:08:18 +00:00
|
|
|
|
|
|
|
/*
|
2016-07-28 22:46:32 +00:00
|
|
|
* Only run node reclaim on the local node or on nodes that do not
|
2006-03-22 08:08:18 +00:00
|
|
|
* have associated processors. This will favor the local processor
|
|
|
|
* over remote processors and spread off node memory allocations
|
|
|
|
* as wide as possible.
|
|
|
|
*/
|
2016-07-28 22:46:32 +00:00
|
|
|
if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
|
|
|
|
return NODE_RECLAIM_NOSCAN;
|
2007-10-17 06:26:01 +00:00
|
|
|
|
2016-07-28 22:46:32 +00:00
|
|
|
if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
|
|
|
|
return NODE_RECLAIM_NOSCAN;
|
2009-06-16 22:33:22 +00:00
|
|
|
|
2016-07-28 22:46:32 +00:00
|
|
|
ret = __node_reclaim(pgdat, gfp_mask, order);
|
|
|
|
clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
|
2007-10-17 06:26:01 +00:00
|
|
|
|
2009-06-16 22:33:23 +00:00
|
|
|
if (!ret)
|
|
|
|
count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
|
|
|
|
|
2007-10-17 06:26:01 +00:00
|
|
|
return ret;
|
2006-03-22 08:08:18 +00:00
|
|
|
}
|
2006-01-19 01:42:31 +00:00
|
|
|
#endif
|
Unevictable LRU Infrastructure
When the system contains lots of mlocked or otherwise unevictable pages,
the pageout code (kswapd) can spend lots of time scanning over these
pages. Worse still, the presence of lots of unevictable pages can confuse
kswapd into thinking that more aggressive pageout modes are required,
resulting in all kinds of bad behaviour.
Infrastructure to manage pages excluded from reclaim--i.e., hidden from
vmscan. Based on a patch by Larry Woodman of Red Hat. Reworked to
maintain "unevictable" pages on a separate per-zone LRU list, to "hide"
them from vmscan.
Kosaki Motohiro added the support for the memory controller unevictable
lru list.
Pages on the unevictable list have both PG_unevictable and PG_lru set.
Thus, PG_unevictable is analogous to and mutually exclusive with
PG_active--it specifies which LRU list the page is on.
The unevictable infrastructure is enabled by a new mm Kconfig option
[CONFIG_]UNEVICTABLE_LRU.
A new function 'page_evictable(page, vma)' in vmscan.c tests whether or
not a page may be evictable. Subsequent patches will add the various
!evictable tests. We'll want to keep these tests light-weight for use in
shrink_active_list() and, possibly, the fault path.
To avoid races between tasks putting pages [back] onto an LRU list and
tasks that might be moving the page from non-evictable to evictable state,
the new function 'putback_lru_page()' -- inverse to 'isolate_lru_page()'
-- tests the "evictability" of a page after placing it on the LRU, before
dropping the reference. If the page has become unevictable,
putback_lru_page() will redo the 'putback', thus moving the page to the
unevictable list. This way, we avoid "stranding" evictable pages on the
unevictable list.
[akpm@linux-foundation.org: fix fallout from out-of-order merge]
[riel@redhat.com: fix UNEVICTABLE_LRU and !PROC_PAGE_MONITOR build]
[nishimura@mxp.nes.nec.co.jp: remove redundant mapping check]
[kosaki.motohiro@jp.fujitsu.com: unevictable-lru-infrastructure: putback_lru_page()/unevictable page handling rework]
[kosaki.motohiro@jp.fujitsu.com: kill unnecessary lock_page() in vmscan.c]
[kosaki.motohiro@jp.fujitsu.com: revert migration change of unevictable lru infrastructure]
[kosaki.motohiro@jp.fujitsu.com: revert to unevictable-lru-infrastructure-kconfig-fix.patch]
[kosaki.motohiro@jp.fujitsu.com: restore patch failure of vmstat-unevictable-and-mlocked-pages-vm-events.patch]
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Debugged-by: Benjamin Kidwell <benjkidwell@yahoo.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 03:26:39 +00:00
|
|
|
|
2022-06-04 21:39:09 +00:00
|
|
|
void check_move_unevictable_pages(struct pagevec *pvec)
|
|
|
|
{
|
|
|
|
struct folio_batch fbatch;
|
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
folio_batch_init(&fbatch);
|
|
|
|
for (i = 0; i < pvec->nr; i++) {
|
|
|
|
struct page *page = pvec->pages[i];
|
|
|
|
|
|
|
|
if (PageTransTail(page))
|
|
|
|
continue;
|
|
|
|
folio_batch_add(&fbatch, page_folio(page));
|
|
|
|
}
|
|
|
|
check_move_unevictable_folios(&fbatch);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
|
|
|
|
|
2008-10-19 03:26:43 +00:00
|
|
|
/**
|
2022-06-04 21:39:09 +00:00
|
|
|
* check_move_unevictable_folios - Move evictable folios to appropriate zone
|
|
|
|
* lru list
|
|
|
|
* @fbatch: Batch of lru folios to check.
|
2008-10-19 03:26:43 +00:00
|
|
|
*
|
2022-06-04 21:39:09 +00:00
|
|
|
* Checks folios for evictability, if an evictable folio is in the unevictable
|
2018-11-06 13:23:24 +00:00
|
|
|
* lru list, moves it to the appropriate evictable lru list. This function
|
2022-06-04 21:39:09 +00:00
|
|
|
* should be only used for lru folios.
|
2008-10-19 03:26:43 +00:00
|
|
|
*/
|
2022-06-04 21:39:09 +00:00
|
|
|
void check_move_unevictable_folios(struct folio_batch *fbatch)
|
2008-10-19 03:26:43 +00:00
|
|
|
{
|
2020-12-15 20:34:29 +00:00
|
|
|
struct lruvec *lruvec = NULL;
|
SHM_UNLOCK: fix Unevictable pages stranded after swap
Commit cc39c6a9bbde ("mm: account skipped entries to avoid looping in
find_get_pages") correctly fixed an infinite loop; but left a problem
that find_get_pages() on shmem would return 0 (appearing to callers to
mean end of tree) when it meets a run of nr_pages swap entries.
The only uses of find_get_pages() on shmem are via pagevec_lookup(),
called from invalidate_mapping_pages(), and from shmctl SHM_UNLOCK's
scan_mapping_unevictable_pages(). The first is already commented, and
not worth worrying about; but the second can leave pages on the
Unevictable list after an unusual sequence of swapping and locking.
Fix that by using shmem_find_get_pages_and_swap() (then ignoring the
swap) instead of pagevec_lookup().
But I don't want to contaminate vmscan.c with shmem internals, nor
shmem.c with LRU locking. So move scan_mapping_unevictable_pages() into
shmem.c, renaming it shmem_unlock_mapping(); and rename
check_move_unevictable_page() to check_move_unevictable_pages(), looping
down an array of pages, oftentimes under the same lock.
Leave out the "rotate unevictable list" block: that's a leftover from
when this was used for /proc/sys/vm/scan_unevictable_pages, whose flawed
handling involved looking at pages at tail of LRU.
Was there significance to the sequence first ClearPageUnevictable, then
test page_evictable, then SetPageUnevictable here? I think not, we're
under LRU lock, and have no barriers between those.
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: <stable@vger.kernel.org> [back to 3.1 but will need respins]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-20 22:34:21 +00:00
|
|
|
int pgscanned = 0;
|
|
|
|
int pgrescued = 0;
|
|
|
|
int i;
|
2008-10-19 03:26:43 +00:00
|
|
|
|
2022-06-04 21:39:09 +00:00
|
|
|
for (i = 0; i < fbatch->nr; i++) {
|
|
|
|
struct folio *folio = fbatch->folios[i];
|
|
|
|
int nr_pages = folio_nr_pages(folio);
|
2020-09-19 04:20:12 +00:00
|
|
|
|
|
|
|
pgscanned += nr_pages;
|
2008-10-19 03:26:43 +00:00
|
|
|
|
2022-06-04 21:39:09 +00:00
|
|
|
/* block memcg migration while the folio moves between lrus */
|
|
|
|
if (!folio_test_clear_lru(folio))
|
mm/lru: introduce TestClearPageLRU()
Currently lru_lock still guards both lru list and page's lru bit, that's
ok. but if we want to use specific lruvec lock on the page, we need to
pin down the page's lruvec/memcg during locking. Just taking lruvec lock
first may be undermined by the page's memcg charge/migration. To fix this
problem, we will clear the lru bit out of locking and use it as pin down
action to block the page isolation in memcg changing.
So now a standard steps of page isolation is following:
1, get_page(); #pin the page avoid to be free
2, TestClearPageLRU(); #block other isolation like memcg change
3, spin_lock on lru_lock; #serialize lru list access
4, delete page from lru list;
This patch start with the first part: TestClearPageLRU, which combines
PageLRU check and ClearPageLRU into a macro func TestClearPageLRU. This
function will be used as page isolation precondition to prevent other
isolations some where else. Then there are may !PageLRU page on lru list,
need to remove BUG() checking accordingly.
There 2 rules for lru bit now:
1, the lru bit still indicate if a page on lru list, just in some
temporary moment(isolating), the page may have no lru bit when
it's on lru list. but the page still must be on lru list when the
lru bit set.
2, have to remove lru bit before delete it from lru list.
As Andrew Morton mentioned this change would dirty cacheline for a page
which isn't on the LRU. But the loss would be acceptable in Rong Chen
<rong.a.chen@intel.com> report:
https://lore.kernel.org/lkml/20200304090301.GB5972@shao2-debian/
Link: https://lkml.kernel.org/r/1604566549-62481-15-git-send-email-alex.shi@linux.alibaba.com
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 20:34:16 +00:00
|
|
|
continue;
|
|
|
|
|
2021-06-30 02:27:31 +00:00
|
|
|
lruvec = folio_lruvec_relock_irq(folio, lruvec);
|
2022-06-04 21:39:09 +00:00
|
|
|
if (folio_evictable(folio) && folio_test_unevictable(folio)) {
|
|
|
|
lruvec_del_folio(lruvec, folio);
|
|
|
|
folio_clear_unevictable(folio);
|
|
|
|
lruvec_add_folio(lruvec, folio);
|
2020-09-19 04:20:12 +00:00
|
|
|
pgrescued += nr_pages;
|
2008-10-19 03:26:43 +00:00
|
|
|
}
|
2022-06-04 21:39:09 +00:00
|
|
|
folio_set_lru(folio);
|
SHM_UNLOCK: fix Unevictable pages stranded after swap
Commit cc39c6a9bbde ("mm: account skipped entries to avoid looping in
find_get_pages") correctly fixed an infinite loop; but left a problem
that find_get_pages() on shmem would return 0 (appearing to callers to
mean end of tree) when it meets a run of nr_pages swap entries.
The only uses of find_get_pages() on shmem are via pagevec_lookup(),
called from invalidate_mapping_pages(), and from shmctl SHM_UNLOCK's
scan_mapping_unevictable_pages(). The first is already commented, and
not worth worrying about; but the second can leave pages on the
Unevictable list after an unusual sequence of swapping and locking.
Fix that by using shmem_find_get_pages_and_swap() (then ignoring the
swap) instead of pagevec_lookup().
But I don't want to contaminate vmscan.c with shmem internals, nor
shmem.c with LRU locking. So move scan_mapping_unevictable_pages() into
shmem.c, renaming it shmem_unlock_mapping(); and rename
check_move_unevictable_page() to check_move_unevictable_pages(), looping
down an array of pages, oftentimes under the same lock.
Leave out the "rotate unevictable list" block: that's a leftover from
when this was used for /proc/sys/vm/scan_unevictable_pages, whose flawed
handling involved looking at pages at tail of LRU.
Was there significance to the sequence first ClearPageUnevictable, then
test page_evictable, then SetPageUnevictable here? I think not, we're
under LRU lock, and have no barriers between those.
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: <stable@vger.kernel.org> [back to 3.1 but will need respins]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-20 22:34:21 +00:00
|
|
|
}
|
2008-10-19 03:26:43 +00:00
|
|
|
|
2020-12-15 20:34:29 +00:00
|
|
|
if (lruvec) {
|
SHM_UNLOCK: fix Unevictable pages stranded after swap
Commit cc39c6a9bbde ("mm: account skipped entries to avoid looping in
find_get_pages") correctly fixed an infinite loop; but left a problem
that find_get_pages() on shmem would return 0 (appearing to callers to
mean end of tree) when it meets a run of nr_pages swap entries.
The only uses of find_get_pages() on shmem are via pagevec_lookup(),
called from invalidate_mapping_pages(), and from shmctl SHM_UNLOCK's
scan_mapping_unevictable_pages(). The first is already commented, and
not worth worrying about; but the second can leave pages on the
Unevictable list after an unusual sequence of swapping and locking.
Fix that by using shmem_find_get_pages_and_swap() (then ignoring the
swap) instead of pagevec_lookup().
But I don't want to contaminate vmscan.c with shmem internals, nor
shmem.c with LRU locking. So move scan_mapping_unevictable_pages() into
shmem.c, renaming it shmem_unlock_mapping(); and rename
check_move_unevictable_page() to check_move_unevictable_pages(), looping
down an array of pages, oftentimes under the same lock.
Leave out the "rotate unevictable list" block: that's a leftover from
when this was used for /proc/sys/vm/scan_unevictable_pages, whose flawed
handling involved looking at pages at tail of LRU.
Was there significance to the sequence first ClearPageUnevictable, then
test page_evictable, then SetPageUnevictable here? I think not, we're
under LRU lock, and have no barriers between those.
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Cc: <stable@vger.kernel.org> [back to 3.1 but will need respins]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-20 22:34:21 +00:00
|
|
|
__count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
|
|
|
|
__count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
|
2020-12-15 20:34:29 +00:00
|
|
|
unlock_page_lruvec_irq(lruvec);
|
mm/lru: introduce TestClearPageLRU()
Currently lru_lock still guards both lru list and page's lru bit, that's
ok. but if we want to use specific lruvec lock on the page, we need to
pin down the page's lruvec/memcg during locking. Just taking lruvec lock
first may be undermined by the page's memcg charge/migration. To fix this
problem, we will clear the lru bit out of locking and use it as pin down
action to block the page isolation in memcg changing.
So now a standard steps of page isolation is following:
1, get_page(); #pin the page avoid to be free
2, TestClearPageLRU(); #block other isolation like memcg change
3, spin_lock on lru_lock; #serialize lru list access
4, delete page from lru list;
This patch start with the first part: TestClearPageLRU, which combines
PageLRU check and ClearPageLRU into a macro func TestClearPageLRU. This
function will be used as page isolation precondition to prevent other
isolations some where else. Then there are may !PageLRU page on lru list,
need to remove BUG() checking accordingly.
There 2 rules for lru bit now:
1, the lru bit still indicate if a page on lru list, just in some
temporary moment(isolating), the page may have no lru bit when
it's on lru list. but the page still must be on lru list when the
lru bit set.
2, have to remove lru bit before delete it from lru list.
As Andrew Morton mentioned this change would dirty cacheline for a page
which isn't on the LRU. But the loss would be acceptable in Rong Chen
<rong.a.chen@intel.com> report:
https://lore.kernel.org/lkml/20200304090301.GB5972@shao2-debian/
Link: https://lkml.kernel.org/r/1604566549-62481-15-git-send-email-alex.shi@linux.alibaba.com
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 20:34:16 +00:00
|
|
|
} else if (pgscanned) {
|
|
|
|
count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
|
2008-10-19 03:26:43 +00:00
|
|
|
}
|
|
|
|
}
|
2022-06-04 21:39:09 +00:00
|
|
|
EXPORT_SYMBOL_GPL(check_move_unevictable_folios);
|