As proposed by Chris, Dave and Jan, don't start foreground writeback IO
inside balance_dirty_pages(). Instead, simply let it idle sleep for some
time to throttle the dirtying task. In the mean while, kick off the
per-bdi flusher thread to do background writeback IO.
RATIONALS
=========
- disk seeks on concurrent writeback of multiple inodes (Dave Chinner)
If every thread doing writes and being throttled start foreground
writeback, it leads to N IO submitters from at least N different
inodes at the same time, end up with N different sets of IO being
issued with potentially zero locality to each other, resulting in
much lower elevator sort/merge efficiency and hence we seek the disk
all over the place to service the different sets of IO.
OTOH, if there is only one submission thread, it doesn't jump between
inodes in the same way when congestion clears - it keeps writing to
the same inode, resulting in large related chunks of sequential IOs
being issued to the disk. This is more efficient than the above
foreground writeback because the elevator works better and the disk
seeks less.
- lock contention and cache bouncing on concurrent IO submitters (Dave Chinner)
With this patchset, the fs_mark benchmark on a 12-drive software RAID0 goes
from CPU bound to IO bound, freeing "3-4 CPUs worth of spinlock contention".
* "CPU usage has dropped by ~55%", "it certainly appears that most of
the CPU time saving comes from the removal of contention on the
inode_wb_list_lock" (IMHO at least 10% comes from the reduction of
cacheline bouncing, because the new code is able to call much less
frequently into balance_dirty_pages() and hence access the global
page states)
* the user space "App overhead" is reduced by 20%, by avoiding the
cacheline pollution by the complex writeback code path
* "for a ~5% throughput reduction", "the number of write IOs have
dropped by ~25%", and the elapsed time reduced from 41:42.17 to
40:53.23.
* On a simple test of 100 dd, it reduces the CPU %system time from 30% to 3%,
and improves IO throughput from 38MB/s to 42MB/s.
- IO size too small for fast arrays and too large for slow USB sticks
The write_chunk used by current balance_dirty_pages() cannot be
directly set to some large value (eg. 128MB) for better IO efficiency.
Because it could lead to more than 1 second user perceivable stalls.
Even the current 4MB write size may be too large for slow USB sticks.
The fact that balance_dirty_pages() starts IO on itself couples the
IO size to wait time, which makes it hard to do suitable IO size while
keeping the wait time under control.
Now it's possible to increase writeback chunk size proportional to the
disk bandwidth. In a simple test of 50 dd's on XFS, 1-HDD, 3GB ram,
the larger writeback size dramatically reduces the seek count to 1/10
(far beyond my expectation) and improves the write throughput by 24%.
- long block time in balance_dirty_pages() hurts desktop responsiveness
Many of us may have the experience: it often takes a couple of seconds
or even long time to stop a heavy writing dd/cp/tar command with
Ctrl-C or "kill -9".
- IO pipeline broken by bumpy write() progress
There are a broad class of "loop {read(buf); write(buf);}" applications
whose read() pipeline will be under-utilized or even come to a stop if
the write()s have long latencies _or_ don't progress in a constant rate.
The current threshold based throttling inherently transfers the large
low level IO completion fluctuations to bumpy application write()s,
and further deteriorates with increasing number of dirtiers and/or bdi's.
For example, when doing 50 dd's + 1 remote rsync to an XFS partition,
the rsync progresses very bumpy in legacy kernel, and throughput is
improved by 67% by this patchset. (plus the larger write chunk size,
it will be 93% speedup).
The new rate based throttling can support 1000+ dd's with excellent
smoothness, low latency and low overheads.
For the above reasons, it's much better to do IO-less and low latency
pauses in balance_dirty_pages().
Jan Kara, Dave Chinner and me explored the scheme to let
balance_dirty_pages() wait for enough writeback IO completions to
safeguard the dirty limit. However it's found to have two problems:
- in large NUMA systems, the per-cpu counters may have big accounting
errors, leading to big throttle wait time and jitters.
- NFS may kill large amount of unstable pages with one single COMMIT.
Because NFS server serves COMMIT with expensive fsync() IOs, it is
desirable to delay and reduce the number of COMMITs. So it's not
likely to optimize away such kind of bursty IO completions, and the
resulted large (and tiny) stall times in IO completion based throttling.
So here is a pause time oriented approach, which tries to control the
pause time in each balance_dirty_pages() invocations, by controlling
the number of pages dirtied before calling balance_dirty_pages(), for
smooth and efficient dirty throttling:
- avoid useless (eg. zero pause time) balance_dirty_pages() calls
- avoid too small pause time (less than 4ms, which burns CPU power)
- avoid too large pause time (more than 200ms, which hurts responsiveness)
- avoid big fluctuations of pause times
It can control pause times at will. The default policy (in a followup
patch) will be to do ~10ms pauses in 1-dd case, and increase to ~100ms
in 1000-dd case.
BEHAVIOR CHANGE
===============
(1) dirty threshold
Users will notice that the applications will get throttled once crossing
the global (background + dirty)/2=15% threshold, and then balanced around
17.5%. Before patch, the behavior is to just throttle it at 20% dirtyable
memory in 1-dd case.
Since the task will be soft throttled earlier than before, it may be
perceived by end users as performance "slow down" if his application
happens to dirty more than 15% dirtyable memory.
(2) smoothness/responsiveness
Users will notice a more responsive system during heavy writeback.
"killall dd" will take effect instantly.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Add two fields to task_struct.
1) account dirtied pages in the individual tasks, for accuracy
2) per-task balance_dirty_pages() call intervals, for flexibility
The balance_dirty_pages() call interval (ie. nr_dirtied_pause) will
scale near-sqrt to the safety gap between dirty pages and threshold.
The main problem of per-task nr_dirtied is, if 1k+ tasks start dirtying
pages at exactly the same time, each task will be assigned a large
initial nr_dirtied_pause, so that the dirty threshold will be exceeded
long before each task reached its nr_dirtied_pause and hence call
balance_dirty_pages().
The solution is to watch for the number of pages dirtied on each CPU in
between the calls into balance_dirty_pages(). If it exceeds ratelimit_pages
(3% dirty threshold), force call balance_dirty_pages() for a chance to
set bdi->dirty_exceeded. In normal situations, this safeguarding
condition is not expected to trigger at all.
On the sqrt in dirty_poll_interval():
It will serve as an initial guess when dirty pages are still in the
freerun area.
When dirty pages are floating inside the dirty control scope [freerun,
limit], a followup patch will use some refined dirty poll interval to
get the desired pause time.
thresh-dirty (MB) sqrt
1 16
2 22
4 32
8 45
16 64
32 90
64 128
128 181
256 256
512 362
1024 512
The above table means, given 1MB (or 1GB) gap and the dd tasks polling
balance_dirty_pages() on every 16 (or 512) pages, the dirty limit won't
be exceeded as long as there are less than 16 (or 512) concurrent dd's.
So sqrt naturally leads to less overheads and more safe concurrent tasks
for large memory servers, which have large (thresh-freerun) gaps.
peter: keep the per-CPU ratelimit for safeguarding the 1k+ tasks case
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: Andrea Righi <andrea@betterlinux.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
There are some imperfections in balanced_dirty_ratelimit.
1) large fluctuations
The dirty_rate used for computing balanced_dirty_ratelimit is merely
averaged in the past 200ms (very small comparing to the 3s estimation
period for write_bw), which makes rather dispersed distribution of
balanced_dirty_ratelimit.
It's pretty hard to average out the singular points by increasing the
estimation period. Considering that the averaging technique will
introduce very undesirable time lags, I give it up totally. (btw, the 3s
write_bw averaging time lag is much more acceptable because its impact
is one-way and therefore won't lead to oscillations.)
The more practical way is filtering -- most singular
balanced_dirty_ratelimit points can be filtered out by remembering some
prev_balanced_rate and prev_prev_balanced_rate. However the more
reliable way is to guard balanced_dirty_ratelimit with task_ratelimit.
2) due to truncates and fs redirties, the (write_bw <=> dirty_rate)
match could become unbalanced, which may lead to large systematical
errors in balanced_dirty_ratelimit. The truncates, due to its possibly
bumpy nature, can hardly be compensated smoothly. So let's face it. When
some over-estimated balanced_dirty_ratelimit brings dirty_ratelimit
high, dirty pages will go higher than the setpoint. task_ratelimit will
in turn become lower than dirty_ratelimit. So if we consider both
balanced_dirty_ratelimit and task_ratelimit and update dirty_ratelimit
only when they are on the same side of dirty_ratelimit, the systematical
errors in balanced_dirty_ratelimit won't be able to bring
dirty_ratelimit far away.
The balanced_dirty_ratelimit estimation may also be inaccurate near
@limit or @freerun, however is less an issue.
3) since we ultimately want to
- keep the fluctuations of task ratelimit as small as possible
- keep the dirty pages around the setpoint as long time as possible
the update policy used for (2) also serves the above goals nicely:
if for some reason the dirty pages are high (task_ratelimit < dirty_ratelimit),
and dirty_ratelimit is low (dirty_ratelimit < balanced_dirty_ratelimit),
there is no point to bring up dirty_ratelimit in a hurry only to hurt
both the above two goals.
So, we make use of task_ratelimit to limit the update of dirty_ratelimit
in two ways:
1) avoid changing dirty rate when it's against the position control target
(the adjusted rate will slow down the progress of dirty pages going
back to setpoint).
2) limit the step size. task_ratelimit is changing values step by step,
leaving a consistent trace comparing to the randomly jumping
balanced_dirty_ratelimit. task_ratelimit also has the nice smaller
errors in stable state and typically larger errors when there are big
errors in rate. So it's a pretty good limiting factor for the step
size of dirty_ratelimit.
Note that bdi->dirty_ratelimit is always tracking balanced_dirty_ratelimit.
task_ratelimit is merely used as a limiting factor.
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
It's all about bdi->dirty_ratelimit, which aims to be (write_bw / N)
when there are N dd tasks.
On write() syscall, use bdi->dirty_ratelimit
============================================
balance_dirty_pages(pages_dirtied)
{
task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio();
pause = pages_dirtied / task_ratelimit;
sleep(pause);
}
On every 200ms, update bdi->dirty_ratelimit
===========================================
bdi_update_dirty_ratelimit()
{
task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio();
balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate;
bdi->dirty_ratelimit = balanced_dirty_ratelimit
}
Estimation of balanced bdi->dirty_ratelimit
===========================================
balanced task_ratelimit
-----------------------
balance_dirty_pages() needs to throttle tasks dirtying pages such that
the total amount of dirty pages stays below the specified dirty limit in
order to avoid memory deadlocks. Furthermore we desire fairness in that
tasks get throttled proportionally to the amount of pages they dirty.
IOW we want to throttle tasks such that we match the dirty rate to the
writeout bandwidth, this yields a stable amount of dirty pages:
dirty_rate == write_bw (1)
The fairness requirement gives us:
task_ratelimit = balanced_dirty_ratelimit
== write_bw / N (2)
where N is the number of dd tasks. We don't know N beforehand, but
still can estimate balanced_dirty_ratelimit within 200ms.
Start by throttling each dd task at rate
task_ratelimit = task_ratelimit_0 (3)
(any non-zero initial value is OK)
After 200ms, we measured
dirty_rate = # of pages dirtied by all dd's / 200ms
write_bw = # of pages written to the disk / 200ms
For the aggressive dd dirtiers, the equality holds
dirty_rate == N * task_rate
== N * task_ratelimit_0 (4)
Or
task_ratelimit_0 == dirty_rate / N (5)
Now we conclude that the balanced task ratelimit can be estimated by
write_bw
balanced_dirty_ratelimit = task_ratelimit_0 * ---------- (6)
dirty_rate
Because with (4) and (5) we can get the desired equality (1):
write_bw
balanced_dirty_ratelimit == (dirty_rate / N) * ----------
dirty_rate
== write_bw / N
Then using the balanced task ratelimit we can compute task pause times like:
task_pause = task->nr_dirtied / task_ratelimit
task_ratelimit with position control
------------------------------------
However, while the above gives us means of matching the dirty rate to
the writeout bandwidth, it at best provides us with a stable dirty page
count (assuming a static system). In order to control the dirty page
count such that it is high enough to provide performance, but does not
exceed the specified limit we need another control.
The dirty position control works by extending (2) to
task_ratelimit = balanced_dirty_ratelimit * pos_ratio (7)
where pos_ratio is a negative feedback function that subjects to
1) f(setpoint) = 1.0
2) df/dx < 0
That is, if the dirty pages are ABOVE the setpoint, we throttle each
task a bit more HEAVY than balanced_dirty_ratelimit, so that the dirty
pages are created less fast than they are cleaned, thus DROP to the
setpoints (and the reverse).
Based on (7) and the assumption that both dirty_ratelimit and pos_ratio
remains CONSTANT for the past 200ms, we get
task_ratelimit_0 = balanced_dirty_ratelimit * pos_ratio (8)
Putting (8) into (6), we get the formula used in
bdi_update_dirty_ratelimit():
write_bw
balanced_dirty_ratelimit *= pos_ratio * ---------- (9)
dirty_rate
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
bdi_position_ratio() provides a scale factor to bdi->dirty_ratelimit, so
that the resulted task rate limit can drive the dirty pages back to the
global/bdi setpoints.
Old scheme is,
|
free run area | throttle area
----------------------------------------+---------------------------->
thresh^ dirty pages
New scheme is,
^ task rate limit
|
| *
| *
| *
|[free run] * [smooth throttled]
| *
| *
| *
..bdi->dirty_ratelimit..........*
| . *
| . *
| . *
| . *
| . *
+-------------------------------.-----------------------*------------>
setpoint^ limit^ dirty pages
The slope of the bdi control line should be
1) large enough to pull the dirty pages to setpoint reasonably fast
2) small enough to avoid big fluctuations in the resulted pos_ratio and
hence task ratelimit
Since the fluctuation range of the bdi dirty pages is typically observed
to be within 1-second worth of data, the bdi control line's slope is
selected to be a linear function of bdi write bandwidth, so that it can
adapt to slow/fast storage devices well.
Assume the bdi control line
pos_ratio = 1.0 + k * (dirty - bdi_setpoint)
where k is the negative slope.
If targeting for 12.5% fluctuation range in pos_ratio when dirty pages
are fluctuating in range
[bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2],
we get slope
k = - 1 / (8 * write_bw)
Let pos_ratio(x_intercept) = 0, we get the parameter used in code:
x_intercept = bdi_setpoint + 8 * write_bw
The global/bdi slopes are nicely complementing each other when the
system has only one major bdi (indicated by bdi_thresh ~= thresh):
1) slope of global control line => scaling to the control scope size
2) slope of main bdi control line => scaling to the writeout bandwidth
so that
- in memory tight systems, (1) becomes strong enough to squeeze dirty
pages inside the control scope
- in large memory systems where the "gravity" of (1) for pulling the
dirty pages to setpoint is too weak, (2) can back (1) up and drive
dirty pages to bdi_setpoint ~= setpoint reasonably fast.
Unfortunately in JBOD setups, the fluctuation range of bdi threshold
is related to memory size due to the interferences between disks. In
this case, the bdi slope will be weighted sum of write_bw and bdi_thresh.
Given equations
span = x_intercept - bdi_setpoint
k = df/dx = - 1 / span
and the extremum values
span = bdi_thresh
dx = bdi_thresh
we get
df = - dx / span = - 1.0
That means, when bdi_dirty deviates bdi_thresh up, pos_ratio and hence
task ratelimit will fluctuate by -100%.
peter: use 3rd order polynomial for the global control line
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Jan Kara <jack@suse.cz>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Introduce the BDI_DIRTIED counter. It will be used for estimating the
bdi's dirty bandwidth.
CC: Jan Kara <jack@suse.cz>
CC: Michael Rubin <mrubin@google.com>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
* 'for-linus' of git://git.kernel.dk/linux-block:
floppy: use del_timer_sync() in init cleanup
blk-cgroup: be able to remove the record of unplugged device
block: Don't check QUEUE_FLAG_SAME_COMP in __blk_complete_request
mm: Add comment explaining task state setting in bdi_forker_thread()
mm: Cleanup clearing of BDI_pending bit in bdi_forker_thread()
block: simplify force plug flush code a little bit
block: change force plug flush call order
block: Fix queue_flag update when rq_affinity goes from 2 to 1
block: separate priority boosting from REQ_META
block: remove READ_META and WRITE_META
xen-blkback: fixed indentation and comments
xen-blkback: Don't disconnect backend until state switched to XenbusStateClosed.
The found entries by find_get_pages() could be all swap entries. In
this case we skip the entries, but make sure the skipped entries are
accounted, so we don't keep looping.
Using nr_found > nr_skip to simplify code as suggested by Eric.
Reported-and-tested-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Xen backend drivers (e.g., blkback and netback) would sometimes fail to
map grant pages into the vmalloc address space allocated with
alloc_vm_area(). The GNTTABOP_map_grant_ref would fail because Xen could
not find the page (in the L2 table) containing the PTEs it needed to
update.
(XEN) mm.c:3846:d0 Could not find L1 PTE for address fbb42000
netback and blkback were making the hypercall from a kernel thread where
task->active_mm != &init_mm and alloc_vm_area() was only updating the page
tables for init_mm. The usual method of deferring the update to the page
tables of other processes (i.e., after taking a fault) doesn't work as a
fault cannot occur during the hypercall.
This would work on some systems depending on what else was using vmalloc.
Fix this by reverting ef691947d8 ("vmalloc: remove vmalloc_sync_all()
from alloc_vm_area()") and add a comment to explain why it's needed.
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Cc: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Ian Campbell <Ian.Campbell@citrix.com>
Cc: Keir Fraser <keir.xen@gmail.com>
Cc: <stable@kernel.org> [3.0.x]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Revert the post-3.0 commit 82f9d486e5 ("memcg: add
memory.vmscan_stat").
The implementation of per-memcg reclaim statistics violates how memcg
hierarchies usually behave: hierarchically.
The reclaim statistics are accounted to child memcgs and the parent
hitting the limit, but not to hierarchy levels in between. Usually,
hierarchical statistics are perfectly recursive, with each level
representing the sum of itself and all its children.
Since this exports statistics to userspace, this may lead to confusion
and problems with changing things after the release, so revert it now,
we can try again later.
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ying Han <yinghan@google.com>
Cc: Balbir Singh <bsingharora@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Without swap, anonymous pages are not scanned. As such, they should not
count when considering force-scanning a small target if there is no swap.
Otherwise, targets are not force-scanned even when their effective scan
number is zero and the other conditions--kswapd/memcg--apply.
This fixes 246e87a939 ("memcg: fix get_scan_count() for small
targets").
[akpm@linux-foundation.org: fix comment]
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Cc: Ying Han <yinghan@google.com>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The vmstat_text array is only defined for CONFIG_SYSFS or CONFIG_PROC_FS,
yet it is referenced for per-node vmstat with CONFIG_NUMA:
drivers/built-in.o: In function `node_read_vmstat':
node.c:(.text+0x1106df): undefined reference to `vmstat_text'
Introduced in commit fa25c503df ("mm: per-node vmstat: show proper
vmstats").
Define the array for CONFIG_NUMA as well.
[akpm@linux-foundation.org: remove unneeded ifdefs]
Signed-off-by: David Rientjes <rientjes@google.com>
Reported-by: Cong Wang <amwang@redhat.com>
Acked-by: Randy Dunlap <rdunlap@xenotime.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When compiling mm/mempolicy.c with struct user copy checks the following
warning is shown:
In file included from arch/x86/include/asm/uaccess.h:572,
from include/linux/uaccess.h:5,
from include/linux/highmem.h:7,
from include/linux/pagemap.h:10,
from include/linux/mempolicy.h:70,
from mm/mempolicy.c:68:
In function `copy_from_user',
inlined from `compat_sys_get_mempolicy' at mm/mempolicy.c:1415:
arch/x86/include/asm/uaccess_64.h:64: warning: call to `copy_from_user_overflow' declared with attribute warning: copy_from_user() buffer size is not provably correct
LD mm/built-in.o
Fix this by passing correct buffer size value.
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>
commit 9d8cebd4bc ("mm: fix mbind vma merge problem") didn't really
fix the mbind vma merge problem due to wrong pgoff value passing to
vma_merge(), which made vma_merge() always return NULL.
Before the patch applied, we are getting a result like:
addr = 0x7fa58f00c000
[snip]
7fa58f00c000-7fa58f00d000 rw-p 00000000 00:00 0
7fa58f00d000-7fa58f00e000 rw-p 00000000 00:00 0
7fa58f00e000-7fa58f00f000 rw-p 00000000 00:00 0
here 7fa58f00c000->7fa58f00f000 we get 3 VMAs which are expected to be
merged described as described in commit 9d8cebd.
Re-testing the patched kernel with the reproducer provided in commit
9d8cebd, we get the correct result:
addr = 0x7ffa5aaa2000
[snip]
7ffa5aaa2000-7ffa5aaa6000 rw-p 00000000 00:00 0
7fffd556f000-7fffd5584000 rw-p 00000000 00:00 0 [stack]
Signed-off-by: Caspar Zhang <caspar@casparzhang.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
bdi_forker_thread() clears BDI_pending bit at the end of the main loop.
However clearing of this bit must not be done in some cases which is
handled by calling 'continue' from switch statement. That's kind of
unusual construct and without a good reason so change the function into
more intuitive code flow.
CC: Wu Fengguang <fengguang.wu@intel.com>
CC: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
The slab has just one free object, adding it to partial list head doesn't make
sense. And it can cause lock contentation. For example,
1. CPU takes the slab from partial list
2. fetch an object
3. switch to another slab
4. free an object, then the slab is added to partial list again
In this way n->list_lock will be heavily contended.
In fact, Alex had a hackbench regression. 3.1-rc1 performance drops about 70%
against 3.0. This patch fixes it.
Acked-by: Christoph Lameter <cl@linux.com>
Reported-by: Alex Shi <alex.shi@intel.com>
Signed-off-by: Shaohua Li <shli@kernel.org>
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Signed-off-by: Pekka Enberg <penberg@kernel.org>
Commit 79dfdaccd1 ("memcg: make oom_lock 0 and 1 based rather than
counter") tried to oom lock the hierarchy and roll back upon
encountering an already locked memcg.
The code is confused when it comes to detecting a locked memcg, though,
so it would fail and rollback after locking one memcg and encountering
an unlocked second one.
The result is that oom-locking hierarchies fails unconditionally and
that every oom killer invocation simply goes to sleep on the oom
waitqueue forever. The tasks practically hang forever without anyone
intervening, possibly holding locks that trip up unrelated tasks, too.
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-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>
ZONE_CONGESTED is only cleared in kswapd, but pages can be freed in any
task. It's possible ZONE_CONGESTED isn't cleared in some cases:
1. the zone is already balanced just entering balance_pgdat() for
order-0 because concurrent tasks free memory. In this case, later
check will skip the zone as it's balanced so the flag isn't cleared.
2. high order balance fallbacks to order-0. quote from Mel: At the
end of balance_pgdat(), kswapd uses the following logic;
If reclaiming at high order {
for each zone {
if all_unreclaimable
skip
if watermark is not met
order = 0
loop again
/* watermark is met */
clear congested
}
}
i.e. it clears ZONE_CONGESTED if it the zone is balanced. if not,
it restarts balancing at order-0. However, if the higher zones are
balanced for order-0, kswapd will miss clearing ZONE_CONGESTED as
that only happens after a zone is shrunk. This can mean that
wait_iff_congested() stalls unnecessarily.
This patch makes kswapd clear ZONE_CONGESTED during its initial
highmem->dma scan for zones that are already balanced.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I get the below warning:
BUG: using smp_processor_id() in preemptible [00000000] code: bash/746
caller is native_sched_clock+0x37/0x6e
Pid: 746, comm: bash Tainted: G W 3.0.0+ #254
Call Trace:
[<ffffffff813435c6>] debug_smp_processor_id+0xc2/0xdc
[<ffffffff8104158d>] native_sched_clock+0x37/0x6e
[<ffffffff81116219>] try_to_free_mem_cgroup_pages+0x7d/0x270
[<ffffffff8114f1f8>] mem_cgroup_force_empty+0x24b/0x27a
[<ffffffff8114ff21>] ? sys_close+0x38/0x138
[<ffffffff8114ff21>] ? sys_close+0x38/0x138
[<ffffffff8114f257>] mem_cgroup_force_empty_write+0x17/0x19
[<ffffffff810c72fb>] cgroup_file_write+0xa8/0xba
[<ffffffff811522d2>] vfs_write+0xb3/0x138
[<ffffffff8115241a>] sys_write+0x4a/0x71
[<ffffffff8114ffd9>] ? sys_close+0xf0/0x138
[<ffffffff8176deab>] system_call_fastpath+0x16/0x1b
sched_clock() can't be used with preempt enabled. And we don't need
fast approach to get clock here, so let's use ktime API.
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Tested-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>
Commit d1a05b6973 ("memcg do not try to drain per-cpu caches without
pages") added a drain_local_stock() call to a preemptible section.
The draining task looks up the cpu-local stock twice to set the
draining-flag, then to drain the stock and clear the flag again. If the
task is migrated to a different CPU in between, noone will clear the
flag on the first stock and it will be forever undrainable. Its charge
can not be recovered and the cgroup can not be deleted anymore.
Properly pin the task to the executing CPU while draining stocks.
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com
Acked-by: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Revert the pass-good area introduced in ffd1f609ab ("writeback:
introduce max-pause and pass-good dirty limits") and make the max-pause
area smaller and safe.
This fixes ~30% performance regression in the ext3 data=writeback
fio_mmap_randwrite_64k/fio_mmap_randrw_64k test cases, where there are
12 JBOD disks, on each disk runs 8 concurrent tasks doing reads+writes.
Using deadline scheduler also has a regression, but not that big as CFQ,
so this suggests we have some write starvation.
The test logs show that
- the disks are sometimes under utilized
- global dirty pages sometimes rush high to the pass-good area for
several hundred seconds, while in the mean time some bdi dirty pages
drop to very low value (bdi_dirty << bdi_thresh). Then suddenly the
global dirty pages dropped under global dirty threshold and bdi_dirty
rush very high (for example, 2 times higher than bdi_thresh). During
which time balance_dirty_pages() is not called at all.
So the problems are
1) The random writes progress so slow that they break the assumption of
the max-pause logic that "8 pages per 200ms is typically more than
enough to curb heavy dirtiers".
2) The max-pause logic ignored task_bdi_thresh and thus opens the possibility
for some bdi's to over dirty pages, leading to (bdi_dirty >> bdi_thresh)
and then (bdi_thresh >> bdi_dirty) for others.
3) The higher max-pause/pass-good thresholds somehow leads to the bad
swing of dirty pages.
The fix is to allow the task to slightly dirty over task_bdi_thresh, but
no way to exceed bdi_dirty and/or global dirty_thresh.
Tests show that it fixed the JBOD regression completely (both behavior
and performance), while still being able to cut down large pause times
in balance_dirty_pages() for single-disk cases.
Reported-by: Li Shaohua <shaohua.li@intel.com>
Tested-by: Li Shaohua <shaohua.li@intel.com>
Acked-by: Jan Kara <jack@suse.cz>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Followup to 33dd4e0ec9 "mm: make some struct page's const" which missed the
HASHED_PAGE_VIRTUAL case.
Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit db64fe0225 ("mm: rewrite vmap layer") introduced code that does
address calculations under the assumption that VMAP_BLOCK_SIZE is a
power of two. However, this might not be true if CONFIG_NR_CPUS is not
set to a power of two.
Wrong vmap_block index/offset values could lead to memory corruption.
However, this has never been observed in practice (or never been
diagnosed correctly); what caught this was the BUG_ON in vb_alloc() that
checks for inconsistent vmap_block indices.
To fix this, ensure that VMAP_BLOCK_SIZE always is a power of two.
BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=31572
Reported-by: Pavel Kysilka <goldenfish@linuxsoft.cz>
Reported-by: Matias A. Fonzo <selk@dragora.org>
Signed-off-by: Clemens Ladisch <clemens@ladisch.de>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: 2.6.28+ <stable@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit 8521fc50d4.
The patch incorrectly assumes that using atomic FLUSHING_CACHED_CHARGE
bit operations is sufficient but that is not true. Johannes Weiner has
reported a crash during parallel memory cgroup removal:
BUG: unable to handle kernel NULL pointer dereference at 0000000000000018
IP: [<ffffffff81083b70>] css_is_ancestor+0x20/0x70
Oops: 0000 [#1] PREEMPT SMP
Pid: 19677, comm: rmdir Tainted: G W 3.0.0-mm1-00188-gf38d32b #35 ECS MCP61M-M3/MCP61M-M3
RIP: 0010:[<ffffffff81083b70>] css_is_ancestor+0x20/0x70
RSP: 0018:ffff880077b09c88 EFLAGS: 00010202
Process rmdir (pid: 19677, threadinfo ffff880077b08000, task ffff8800781bb310)
Call Trace:
[<ffffffff810feba3>] mem_cgroup_same_or_subtree+0x33/0x40
[<ffffffff810feccf>] drain_all_stock+0x11f/0x170
[<ffffffff81103211>] mem_cgroup_force_empty+0x231/0x6d0
[<ffffffff811036c4>] mem_cgroup_pre_destroy+0x14/0x20
[<ffffffff81080559>] cgroup_rmdir+0xb9/0x500
[<ffffffff81114d26>] vfs_rmdir+0x86/0xe0
[<ffffffff81114e7b>] do_rmdir+0xfb/0x110
[<ffffffff81114ea6>] sys_rmdir+0x16/0x20
[<ffffffff8154d76b>] system_call_fastpath+0x16/0x1b
We are crashing because we try to dereference cached memcg when we are
checking whether we should wait for draining on the cache. The cache is
already cleaned up, though.
There is also a theoretical chance that the cached memcg gets freed
between we test for the FLUSHING_CACHED_CHARGE and dereference it in
mem_cgroup_same_or_subtree:
CPU0 CPU1 CPU2
mem=stock->cached
stock->cached=NULL
clear_bit
test_and_set_bit
test_bit() ...
<preempted> mem_cgroup_destroy
use after free
The percpu_charge_mutex protected from this race because sync draining
is exclusive.
It is safer to revert now and come up with a more parallel
implementation later.
Signed-off-by: Michal Hocko <mhocko@suse.cz>
Reported-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
deactivate_slab() has the comparison if more than the minimum number of
partial pages are in the partial list wrong. An effect of this may be that
empty pages are not freed from deactivate_slab(). The result could be an
OOM due to growth of the partial slabs per node. Frees mostly occur from
__slab_free which is okay so this would only affect use cases where a lot
of switching around of per cpu slabs occur.
Switching per cpu slabs occurs with high frequency if debugging options are
enabled.
Reported-and-tested-by: Xiaotian Feng <xtfeng@gmail.com>
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Pekka Enberg <penberg@kernel.org>
The check_bytes() function is used by slub debugging. It returns a pointer
to the first unmatching byte for a character in the given memory area.
If the character for matching byte is greater than 0x80, check_bytes()
doesn't work. Becuase 64-bit pattern is generated as below.
value64 = value | value << 8 | value << 16 | value << 24;
value64 = value64 | value64 << 32;
The integer promotions are performed and sign-extended as the type of value
is u8. The upper 32 bits of value64 is 0xffffffff in the first line, and
the second line has no effect.
This fixes the 64-bit pattern generation.
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Matt Mackall <mpm@selenic.com>
Reviewed-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Pekka Enberg <penberg@kernel.org>
* 'core-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
slab, lockdep: Annotate the locks before using them
lockdep: Clear whole lockdep_map on initialization
slab, lockdep: Annotate slab -> rcu -> debug_object -> slab
lockdep: Fix up warning
lockdep: Fix trace_hardirqs_on_caller()
futex: Fix regression with read only mappings
Lockdep thinks there's lock recursion through:
kmem_cache_free()
cache_flusharray()
spin_lock(&l3->list_lock) <----------------.
free_block() |
slab_destroy() |
call_rcu() |
debug_object_activate() |
debug_object_init() |
__debug_object_init() |
kmem_cache_alloc() |
cache_alloc_refill() |
spin_lock(&l3->list_lock) --'
Now debug objects doesn't use SLAB_DESTROY_BY_RCU and hence there is no
actual possibility of recursing. Luckily debug objects marks it slab
with SLAB_DEBUG_OBJECTS so we can identify the thing.
Mark all SLAB_DEBUG_OBJECTS (all one!) slab caches with a special
lockdep key so that lockdep sees its a different cachep.
Also add a WARN on trying to create a SLAB_DESTROY_BY_RCU |
SLAB_DEBUG_OBJECTS cache, to avoid possible future trouble.
Reported-and-tested-by: Sebastian Siewior <sebastian@breakpoint.cc>
[ fixes to the initial patch ]
Reported-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Pekka Enberg <penberg@kernel.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1311341165.27400.58.camel@twins
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'apei-release' of git://git.kernel.org/pub/scm/linux/kernel/git/lenb/linux-acpi-2.6:
ACPI, APEI, EINJ Param support is disabled by default
APEI GHES: 32-bit buildfix
ACPI: APEI build fix
ACPI, APEI, GHES: Add hardware memory error recovery support
HWPoison: add memory_failure_queue()
ACPI, APEI, GHES, Error records content based throttle
ACPI, APEI, GHES, printk support for recoverable error via NMI
lib, Make gen_pool memory allocator lockless
lib, Add lock-less NULL terminated single list
Add Kconfig option ARCH_HAVE_NMI_SAFE_CMPXCHG
ACPI, APEI, Add WHEA _OSC support
ACPI, APEI, Add APEI bit support in generic _OSC call
ACPI, APEI, GHES, Support disable GHES at boot time
ACPI, APEI, GHES, Prevent GHES to be built as module
ACPI, APEI, Use apei_exec_run_optional in APEI EINJ and ERST
ACPI, APEI, Add apei_exec_run_optional
ACPI, APEI, GHES, Do not ratelimit fatal error printk before panic
ACPI, APEI, ERST, Fix erst-dbg long record reading issue
ACPI, APEI, ERST, Prevent erst_dbg from loading if ERST is disabled
Make the radix_tree exceptional cases, mostly in filemap.c, clearer.
It's hard to devise a suitable snappy name that illuminates the use by
shmem/tmpfs for swap, while keeping filemap/pagecache/radix_tree
generality. And akpm points out that /* radix_tree_deref_retry(page) */
comments look like calls that have been commented out for unknown
reason.
Skirt the naming difficulty by rearranging these blocks to handle the
transient radix_tree_deref_retry(page) case first; then just explain the
remaining shmem/tmpfs swap case in a comment.
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We have already acknowledged that swapoff of a tmpfs file is slower than
it was before conversion to the generic radix_tree: a little slower
there will be acceptable, if the hotter paths are faster.
But it was a shock to find swapoff of a 500MB file 20 times slower on my
laptop, taking 10 minutes; and at that rate it significantly slows down
my testing.
Now, most of that turned out to be overhead from PROVE_LOCKING and
PROVE_RCU: without those it was only 4 times slower than before; and
more realistic tests on other machines don't fare as badly.
I've tried a number of things to improve it, including tagging the swap
entries, then doing lookup by tag: I'd expected that to halve the time,
but in practice it's erratic, and often counter-productive.
The only change I've so far found to make a consistent improvement, is
to short-circuit the way we go back and forth, gang lookup packing
entries into the array supplied, then shmem scanning that array for the
target entry. Scanning in place doubles the speed, so it's now only
twice as slow as before (or three times slower when the PROVEs are on).
So, add radix_tree_locate_item() as an expedient, once-off,
single-caller hack to do the lookup directly in place. #ifdef it on
CONFIG_SHMEM and CONFIG_SWAP, as much to document its limited
applicability as save space in other configurations. And, sadly,
#include sched.h for cond_resched().
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Remove PageSwapBacked (!page_is_file_cache) cases from
add_to_page_cache_locked() and add_to_page_cache_lru(): those pages now
go through shmem_add_to_page_cache().
Remove a comment on maximum tmpfs size from fsstack_copy_inode_size(),
and add a comment on swap entries to invalidate_mapping_pages().
And mincore_page() uses find_get_page() on what might be shmem or a
tmpfs file: allow for a radix_tree_exceptional_entry(), and proceed to
find_get_page() on swapper_space if so (oh, swapper_space needs #ifdef).
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
But we've not yet removed the old swp_entry_t i_direct[16] from
shmem_inode_info. That's because it was still being shared with the
inline symlink. Remove it now (saving 64 or 128 bytes from shmem inode
size), and use kmemdup() for short symlinks, say, those up to 128 bytes.
I wonder why mpol_free_shared_policy() is done in shmem_destroy_inode()
rather than shmem_evict_inode(), where we usually do such freeing? I
guess it doesn't matter, and I'm not into NUMA mpol testing right now.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Pekka Enberg <penberg@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Convert shmem_writepage() to use shmem_delete_from_page_cache() to use
shmem_radix_tree_replace() to substitute swap entry for page pointer
atomically in the radix tree.
As with shmem_add_to_page_cache(), it's not entirely satisfactory to be
copying such code from delete_from_swap_cache, but again judged easier
to sell than making its other callers go through the extras.
Remove the toy implementation's shmem_put_swap() and shmem_get_swap(),
now unreferenced, and the hack to disable swap: it's now good to go.
The way things have worked out, info->lock no longer helps to guard the
shmem_swaplist: we increment swapped under shmem_swaplist_mutex only.
That global mutex exclusion between shmem_writepage() and shmem_unuse()
is not pretty, and we ought to find another way; but it's been forced on
us by recent race discoveries, not a consequence of this patchset.
And what has become of the WARN_ON_ONCE(1) free_swap_and_cache() if a
swap entry was found already present? That's no longer possible, the
(unknown) one inserting this page into filecache would hit the swap
entry occupying that slot.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Remove mem_cgroup_shmem_charge_fallback(): it was only required when we
had to move swappage to filecache with GFP_NOWAIT.
Remove the GFP_NOWAIT special case from mem_cgroup_cache_charge(), by
moving its call out from shmem_add_to_page_cache() to two of thats three
callers. But leave it doing mem_cgroup_uncharge_cache_page() on error:
although asymmetrical, it's easier for all 3 callers to handle.
These two changes would also be appropriate if anyone were to start
using shmem_read_mapping_page_gfp() with GFP_NOWAIT.
Remove mem_cgroup_get_shmem_target(): mc_handle_file_pte() can test
radix_tree_exceptional_entry() to get what it needs for itself.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Convert shmem_getpage_gfp(), the engine-room of shmem, to expect page or
swap entry returned from radix tree by find_lock_page().
Whereas the repetitive old method proceeded mainly under info->lock,
dropping and repeating whenever one of the conditions needed was not
met, now we can proceed without it, leaving shmem_add_to_page_cache() to
check for a race.
This way there is no need to preallocate a page, no need for an early
radix_tree_preload(), no need for mem_cgroup_shmem_charge_fallback().
Move the error unwinding down to the bottom instead of repeating it
throughout. ENOSPC handling is a little different from before: there is
no longer any race between find_lock_page() and finding swap, but we can
arrive at ENOSPC before calling shmem_recalc_inode(), which might
occasionally discover freed space.
Be stricter to check i_size before returning. info->lock is used for
little but alloced, swapped, i_blocks updates. Move i_blocks updates
out from under the max_blocks check, so even an unlimited size=0 mount
can show accurate du.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Convert shmem_unuse_inode() to use a lockless gang lookup of the radix
tree, searching for matching swap.
This is somewhat slower than the old method: because of repeated radix
tree descents, because of copying entries up, but probably most because
the old method noted and skipped once a vector page was cleared of swap.
Perhaps we can devise a use of radix tree tagging to achieve that later.
shmem_add_to_page_cache() uses shmem_radix_tree_replace() to compensate
for the lockless lookup by checking that the expected entry is in place,
under lock. It is not very satisfactory to be copying this much from
add_to_page_cache_locked(), but I think easier to sell than insisting
that every caller of add_to_page_cache*() go through the extras.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Disable the toy swapping implementation in shmem_writepage() - it's hard
to support two schemes at once - and convert shmem_truncate_range() to a
lockless gang lookup of swap entries along with pages, freeing both.
Since the second loop tightens its noose until all entries of either
kind have been squeezed out (and we shall make sure that there's not an
instant when neither is visible), there is no longer a need for yet
another pass below.
shmem_radix_tree_replace() compensates for the lockless lookup by
checking that the expected entry is in place, under lock, before
replacing it. Here it just deletes, but will be used in later patches
to substitute swap entry for page or page for swap entry.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Bring truncate.c's code for truncate_inode_pages_range() inline into
shmem_truncate_range(), replacing its first call (there's a followup
call below, but leave that one, it will disappear next).
Don't play with it yet, apart from leaving out the cleancache flush, and
(importantly) the nrpages == 0 skip, and moving shmem_setattr()'s
partial page preparation into its partial page handling.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
While it's at its least, make a number of boring nitpicky cleanups to
shmem.c, mostly for consistency of variable naming. Things like "swap"
instead of "entry", "pgoff_t index" instead of "unsigned long idx".
And since everything else here is prefixed "shmem_", better change
init_tmpfs() to shmem_init().
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The maximum size of a shmem/tmpfs file has been limited by the maximum
size of its triple-indirect swap vector. With 4kB page size, maximum
filesize was just over 2TB on a 32-bit kernel, but sadly one eighth of
that on a 64-bit kernel. (With 8kB page size, maximum filesize was just
over 4TB on a 64-bit kernel, but 16TB on a 32-bit kernel,
MAX_LFS_FILESIZE being then more restrictive than swap vector layout.)
It's a shame that tmpfs should be more restrictive than ramfs, and this
limitation has now been noticed. Add another level to the swap vector?
No, it became obscure and hard to maintain, once I complicated it to
make use of highmem pages nine years ago: better choose another way.
Surely, if 2.4 had had the radix tree pagecache introduced in 2.5, then
tmpfs would never have invented its own peculiar radix tree: we would
have fitted swap entries into the common radix tree instead, in much the
same way as we fit swap entries into page tables.
And why should each file have a separate radix tree for its pages and
for its swap entries? The swap entries are required precisely where and
when the pages are not. We want to put them together in a single radix
tree: which can then avoid much of the locking which was needed to
prevent them from being exchanged underneath us.
This also avoids the waste of memory devoted to swap vectors, first in
the shmem_inode itself, then at least two more pages once a file grew
beyond 16 data pages (pages accounted by df and du, but not by memcg).
Allocated upfront, to avoid allocation when under swapping pressure, but
pure waste when CONFIG_SWAP is not set - I have never spattered around
the ifdefs to prevent that, preferring this move to sharing the common
radix tree instead.
There are three downsides to sharing the radix tree. One, that it binds
tmpfs more tightly to the rest of mm, either requiring knowledge of swap
entries in radix tree there, or duplication of its code here in shmem.c.
I believe that the simplications and memory savings (and probable higher
performance, not yet measured) justify that.
Two, that on HIGHMEM systems with SWAP enabled, it's the lowmem radix
nodes that cannot be freed under memory pressure - whereas before it was
the less precious highmem swap vector pages that could not be freed.
I'm hoping that 64-bit has now been accessible for long enough, that the
highmem argument has grown much less persuasive.
Three, that swapoff is slower than it used to be on tmpfs files, since
it's using a simple generic mechanism not tailored to it: I find this
noticeable, and shall want to improve, but maybe nobody else will
notice.
So... now remove most of the old swap vector code from shmem.c. But,
for the moment, keep the simple i_direct vector of 16 pages, with simple
accessors shmem_put_swap() and shmem_get_swap(), as a toy implementation
to help mark where swap needs to be handled in subsequent patches.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If swap entries are to be stored along with struct page pointers in a
radix tree, they need to be distinguished as exceptional entries.
Most of the handling of swap entries in radix tree will be contained in
shmem.c, but a few functions in filemap.c's common code need to check
for their appearance: find_get_page(), find_lock_page(),
find_get_pages() and find_get_pages_contig().
So as not to slow their fast paths, tuck those checks inside the
existing checks for unlikely radix_tree_deref_slot(); except for
find_lock_page(), where it is an added test. And make it a BUG in
find_get_pages_tag(), which is not applied to tmpfs files.
A part of the reason for eliminating shmem_readpage() earlier, was to
minimize the places where common code would need to allow for swap
entries.
The swp_entry_t known to swapfile.c must be massaged into a slightly
different form when stored in the radix tree, just as it gets massaged
into a pte_t when stored in page tables.
In an i386 kernel this limits its information (type and page offset) to
30 bits: given 32 "types" of swapfile and 4kB pagesize, that's a maximum
swapfile size of 128GB. Which is less than the 512GB we previously
allowed with X86_PAE (where the swap entry can occupy the entire upper
32 bits of a pte_t), but not a new limitation on 32-bit without PAE; and
there's not a new limitation on 64-bit (where swap filesize is already
limited to 16TB by a 32-bit page offset). Thirty areas of 128GB is
probably still enough swap for a 64GB 32-bit machine.
Provide swp_to_radix_entry() and radix_to_swp_entry() conversions, and
enforce filesize limit in read_swap_header(), just as for ptes.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
A patchset to extend tmpfs to MAX_LFS_FILESIZE by abandoning its
peculiar swap vector, instead keeping a file's swap entries in the same
radix tree as its struct page pointers: thus saving memory, and
simplifying its code and locking.
This patch:
The radix_tree is used by several subsystems for different purposes. A
major use is to store the struct page pointers of a file's pagecache for
memory management. But what if mm wanted to store something other than
page pointers there too?
The low bit of a radix_tree entry is already used to denote an indirect
pointer, for internal use, and the unlikely radix_tree_deref_retry()
case.
Define the next bit as denoting an exceptional entry, and supply inline
functions radix_tree_exception() to return non-0 in either unlikely
case, and radix_tree_exceptional_entry() to return non-0 in the second
case.
If a subsystem already uses radix_tree with that bit set, no problem: it
does not affect internal workings at all, but is defined for the
convenience of those storing well-aligned pointers in the radix_tree.
The radix_tree_gang_lookups have an implicit assumption that the caller
can deduce the offset of each entry returned e.g. by the page->index of
a struct page. But that may not be feasible for some kinds of item to
be stored there.
radix_tree_gang_lookup_slot() allow for an optional indices argument,
output array in which to return those offsets. The same could be added
to other radix_tree_gang_lookups, but for now keep it to the only one
for which we need it.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
init_fault_attr_dentries() is used to export fault_attr via debugfs.
But it can only export it in debugfs root directory.
Per Forlin is working on mmc_fail_request which adds support to inject
data errors after a completed host transfer in MMC subsystem.
The fault_attr for mmc_fail_request should be defined per mmc host and
export it in debugfs directory per mmc host like
/sys/kernel/debug/mmc0/mmc_fail_request.
init_fault_attr_dentries() doesn't help for mmc_fail_request. So this
introduces fault_create_debugfs_attr() which is able to create a
directory in the arbitrary directory and replace
init_fault_attr_dentries().
[akpm@linux-foundation.org: extraneous semicolon, per Randy]
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Tested-by: Per Forlin <per.forlin@linaro.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Randy Dunlap <rdunlap@xenotime.net>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
