2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2006-10-03 21:01:26 +00:00
|
|
|
* mm/page-writeback.c
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* Copyright (C) 2002, Linus Torvalds.
|
2007-10-17 06:25:50 +00:00
|
|
|
* Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* Contains functions related to writing back dirty pages at the
|
|
|
|
* address_space level.
|
|
|
|
*
|
2008-10-16 05:01:59 +00:00
|
|
|
* 10Apr2002 Andrew Morton
|
2005-04-16 22:20:36 +00:00
|
|
|
* Initial version
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/kernel.h>
|
2011-10-16 06:01:52 +00:00
|
|
|
#include <linux/export.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/spinlock.h>
|
|
|
|
#include <linux/fs.h>
|
|
|
|
#include <linux/mm.h>
|
|
|
|
#include <linux/swap.h>
|
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/pagemap.h>
|
|
|
|
#include <linux/writeback.h>
|
|
|
|
#include <linux/init.h>
|
|
|
|
#include <linux/backing-dev.h>
|
2006-12-10 10:19:27 +00:00
|
|
|
#include <linux/task_io_accounting_ops.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/blkdev.h>
|
|
|
|
#include <linux/mpage.h>
|
2006-09-26 06:30:57 +00:00
|
|
|
#include <linux/rmap.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/percpu.h>
|
|
|
|
#include <linux/notifier.h>
|
|
|
|
#include <linux/smp.h>
|
|
|
|
#include <linux/sysctl.h>
|
|
|
|
#include <linux/cpu.h>
|
|
|
|
#include <linux/syscalls.h>
|
2011-09-16 06:31:11 +00:00
|
|
|
#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
|
2006-08-29 18:06:09 +00:00
|
|
|
#include <linux/pagevec.h>
|
2012-05-24 16:59:11 +00:00
|
|
|
#include <linux/timer.h>
|
2013-02-07 15:47:07 +00:00
|
|
|
#include <linux/sched/rt.h>
|
2013-09-11 21:22:36 +00:00
|
|
|
#include <linux/mm_inline.h>
|
2010-07-07 03:24:07 +00:00
|
|
|
#include <trace/events/writeback.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2013-09-11 21:22:36 +00:00
|
|
|
#include "internal.h"
|
|
|
|
|
2011-06-20 04:18:42 +00:00
|
|
|
/*
|
|
|
|
* Sleep at most 200ms at a time in balance_dirty_pages().
|
|
|
|
*/
|
|
|
|
#define MAX_PAUSE max(HZ/5, 1)
|
|
|
|
|
2011-12-06 19:17:17 +00:00
|
|
|
/*
|
|
|
|
* Try to keep balance_dirty_pages() call intervals higher than this many pages
|
|
|
|
* by raising pause time to max_pause when falls below it.
|
|
|
|
*/
|
|
|
|
#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
|
|
|
|
|
2010-08-29 17:22:30 +00:00
|
|
|
/*
|
|
|
|
* Estimate write bandwidth at 200ms intervals.
|
|
|
|
*/
|
|
|
|
#define BANDWIDTH_INTERVAL max(HZ/5, 1)
|
|
|
|
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
#define RATELIMIT_CALC_SHIFT 10
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
|
|
|
|
* will look to see if it needs to force writeback or throttling.
|
|
|
|
*/
|
|
|
|
static long ratelimit_pages = 32;
|
|
|
|
|
|
|
|
/* The following parameters are exported via /proc/sys/vm */
|
|
|
|
|
|
|
|
/*
|
2009-09-23 17:37:09 +00:00
|
|
|
* Start background writeback (via writeback threads) at this percentage
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2009-03-23 00:57:38 +00:00
|
|
|
int dirty_background_ratio = 10;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
/*
|
|
|
|
* dirty_background_bytes starts at 0 (disabled) so that it is a function of
|
|
|
|
* dirty_background_ratio * the amount of dirtyable memory
|
|
|
|
*/
|
|
|
|
unsigned long dirty_background_bytes;
|
|
|
|
|
2008-02-05 06:29:20 +00:00
|
|
|
/*
|
|
|
|
* free highmem will not be subtracted from the total free memory
|
|
|
|
* for calculating free ratios if vm_highmem_is_dirtyable is true
|
|
|
|
*/
|
|
|
|
int vm_highmem_is_dirtyable;
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* The generator of dirty data starts writeback at this percentage
|
|
|
|
*/
|
2009-03-23 00:57:38 +00:00
|
|
|
int vm_dirty_ratio = 20;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
/*
|
|
|
|
* vm_dirty_bytes starts at 0 (disabled) so that it is a function of
|
|
|
|
* vm_dirty_ratio * the amount of dirtyable memory
|
|
|
|
*/
|
|
|
|
unsigned long vm_dirty_bytes;
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2009-03-31 22:23:18 +00:00
|
|
|
* The interval between `kupdate'-style writebacks
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2009-05-17 05:56:28 +00:00
|
|
|
unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2012-03-22 02:33:00 +00:00
|
|
|
EXPORT_SYMBOL_GPL(dirty_writeback_interval);
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2009-03-31 22:23:18 +00:00
|
|
|
* The longest time for which data is allowed to remain dirty
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2009-05-17 05:56:28 +00:00
|
|
|
unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Flag that makes the machine dump writes/reads and block dirtyings.
|
|
|
|
*/
|
|
|
|
int block_dump;
|
|
|
|
|
|
|
|
/*
|
2006-03-24 11:15:49 +00:00
|
|
|
* Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
|
|
|
|
* a full sync is triggered after this time elapses without any disk activity.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
|
|
|
int laptop_mode;
|
|
|
|
|
|
|
|
EXPORT_SYMBOL(laptop_mode);
|
|
|
|
|
|
|
|
/* End of sysctl-exported parameters */
|
|
|
|
|
2011-03-02 21:54:09 +00:00
|
|
|
unsigned long global_dirty_limit;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2007-10-17 06:25:50 +00:00
|
|
|
/*
|
|
|
|
* Scale the writeback cache size proportional to the relative writeout speeds.
|
|
|
|
*
|
|
|
|
* We do this by keeping a floating proportion between BDIs, based on page
|
|
|
|
* writeback completions [end_page_writeback()]. Those devices that write out
|
|
|
|
* pages fastest will get the larger share, while the slower will get a smaller
|
|
|
|
* share.
|
|
|
|
*
|
|
|
|
* We use page writeout completions because we are interested in getting rid of
|
|
|
|
* dirty pages. Having them written out is the primary goal.
|
|
|
|
*
|
|
|
|
* We introduce a concept of time, a period over which we measure these events,
|
|
|
|
* because demand can/will vary over time. The length of this period itself is
|
|
|
|
* measured in page writeback completions.
|
|
|
|
*
|
|
|
|
*/
|
2012-05-24 16:59:11 +00:00
|
|
|
static struct fprop_global writeout_completions;
|
|
|
|
|
|
|
|
static void writeout_period(unsigned long t);
|
|
|
|
/* Timer for aging of writeout_completions */
|
|
|
|
static struct timer_list writeout_period_timer =
|
|
|
|
TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
|
|
|
|
static unsigned long writeout_period_time = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Length of period for aging writeout fractions of bdis. This is an
|
|
|
|
* arbitrarily chosen number. The longer the period, the slower fractions will
|
|
|
|
* reflect changes in current writeout rate.
|
|
|
|
*/
|
|
|
|
#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
|
2007-10-17 06:25:50 +00:00
|
|
|
|
mm: try to distribute dirty pages fairly across zones
The maximum number of dirty pages that exist in the system at any time is
determined by a number of pages considered dirtyable and a user-configured
percentage of those, or an absolute number in bytes.
This number of dirtyable pages is the sum of memory provided by all the
zones in the system minus their lowmem reserves and high watermarks, so
that the system can retain a healthy number of free pages without having
to reclaim dirty pages.
But there is a flaw in that we have a zoned page allocator which does not
care about the global state but rather the state of individual memory
zones. And right now there is nothing that prevents one zone from filling
up with dirty pages while other zones are spared, which frequently leads
to situations where kswapd, in order to restore the watermark of free
pages, does indeed have to write pages from that zone's LRU list. This
can interfere so badly with IO from the flusher threads that major
filesystems (btrfs, xfs, ext4) mostly ignore write requests from reclaim
already, taking away the VM's only possibility to keep such a zone
balanced, aside from hoping the flushers will soon clean pages from that
zone.
Enter per-zone dirty limits. They are to a zone's dirtyable memory what
the global limit is to the global amount of dirtyable memory, and try to
make sure that no single zone receives more than its fair share of the
globally allowed dirty pages in the first place. As the number of pages
considered dirtyable excludes the zones' lowmem reserves and high
watermarks, the maximum number of dirty pages in a zone is such that the
zone can always be balanced without requiring page cleaning.
As this is a placement decision in the page allocator and pages are
dirtied only after the allocation, this patch allows allocators to pass
__GFP_WRITE when they know in advance that the page will be written to and
become dirty soon. The page allocator will then attempt to allocate from
the first zone of the zonelist - which on NUMA is determined by the task's
NUMA memory policy - that has not exceeded its dirty limit.
At first glance, it would appear that the diversion to lower zones can
increase pressure on them, but this is not the case. With a full high
zone, allocations will be diverted to lower zones eventually, so it is
more of a shift in timing of the lower zone allocations. Workloads that
previously could fit their dirty pages completely in the higher zone may
be forced to allocate from lower zones, but the amount of pages that
"spill over" are limited themselves by the lower zones' dirty constraints,
and thus unlikely to become a problem.
For now, the problem of unfair dirty page distribution remains for NUMA
configurations where the zones allowed for allocation are in sum not big
enough to trigger the global dirty limits, wake up the flusher threads and
remedy the situation. Because of this, an allocation that could not
succeed on any of the considered zones is allowed to ignore the dirty
limits before going into direct reclaim or even failing the allocation,
until a future patch changes the global dirty throttling and flusher
thread activation so that they take individual zone states into account.
Test results
15M DMA + 3246M DMA32 + 504 Normal = 3765M memory
40% dirty ratio
16G USB thumb drive
10 runs of dd if=/dev/zero of=disk/zeroes bs=32k count=$((10 << 15))
seconds nr_vmscan_write
(stddev) min| median| max
xfs
vanilla: 549.747( 3.492) 0.000| 0.000| 0.000
patched: 550.996( 3.802) 0.000| 0.000| 0.000
fuse-ntfs
vanilla: 1183.094(53.178) 54349.000| 59341.000| 65163.000
patched: 558.049(17.914) 0.000| 0.000| 43.000
btrfs
vanilla: 573.679(14.015) 156657.000| 460178.000| 606926.000
patched: 563.365(11.368) 0.000| 0.000| 1362.000
ext4
vanilla: 561.197(15.782) 0.000|2725438.000|4143837.000
patched: 568.806(17.496) 0.000| 0.000| 0.000
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 23:07:49 +00:00
|
|
|
/*
|
|
|
|
* In a memory zone, there is a certain amount of pages we consider
|
|
|
|
* available for the page cache, which is essentially the number of
|
|
|
|
* free and reclaimable pages, minus some zone reserves to protect
|
|
|
|
* lowmem and the ability to uphold the zone's watermarks without
|
|
|
|
* requiring writeback.
|
|
|
|
*
|
|
|
|
* This number of dirtyable pages is the base value of which the
|
|
|
|
* user-configurable dirty ratio is the effictive number of pages that
|
|
|
|
* are allowed to be actually dirtied. Per individual zone, or
|
|
|
|
* globally by using the sum of dirtyable pages over all zones.
|
|
|
|
*
|
|
|
|
* Because the user is allowed to specify the dirty limit globally as
|
|
|
|
* absolute number of bytes, calculating the per-zone dirty limit can
|
|
|
|
* require translating the configured limit into a percentage of
|
|
|
|
* global dirtyable memory first.
|
|
|
|
*/
|
|
|
|
|
2014-01-29 22:05:39 +00:00
|
|
|
/**
|
|
|
|
* zone_dirtyable_memory - number of dirtyable pages in a zone
|
|
|
|
* @zone: the zone
|
|
|
|
*
|
|
|
|
* Returns the zone's number of pages potentially available for dirty
|
|
|
|
* page cache. This is the base value for the per-zone dirty limits.
|
|
|
|
*/
|
|
|
|
static unsigned long zone_dirtyable_memory(struct zone *zone)
|
|
|
|
{
|
|
|
|
unsigned long nr_pages;
|
|
|
|
|
|
|
|
nr_pages = zone_page_state(zone, NR_FREE_PAGES);
|
|
|
|
nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
|
|
|
|
|
2014-01-29 22:05:41 +00:00
|
|
|
nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
|
|
|
|
nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
|
2014-01-29 22:05:39 +00:00
|
|
|
|
|
|
|
return nr_pages;
|
|
|
|
}
|
|
|
|
|
2012-01-10 23:06:57 +00:00
|
|
|
static unsigned long highmem_dirtyable_memory(unsigned long total)
|
|
|
|
{
|
|
|
|
#ifdef CONFIG_HIGHMEM
|
|
|
|
int node;
|
|
|
|
unsigned long x = 0;
|
|
|
|
|
|
|
|
for_each_node_state(node, N_HIGH_MEMORY) {
|
2014-01-29 22:05:39 +00:00
|
|
|
struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
|
2012-01-10 23:06:57 +00:00
|
|
|
|
2014-01-29 22:05:39 +00:00
|
|
|
x += zone_dirtyable_memory(z);
|
2012-01-10 23:06:57 +00:00
|
|
|
}
|
2012-12-20 23:05:07 +00:00
|
|
|
/*
|
|
|
|
* Unreclaimable memory (kernel memory or anonymous memory
|
|
|
|
* without swap) can bring down the dirtyable pages below
|
|
|
|
* the zone's dirty balance reserve and the above calculation
|
|
|
|
* will underflow. However we still want to add in nodes
|
|
|
|
* which are below threshold (negative values) to get a more
|
|
|
|
* accurate calculation but make sure that the total never
|
|
|
|
* underflows.
|
|
|
|
*/
|
|
|
|
if ((long)x < 0)
|
|
|
|
x = 0;
|
|
|
|
|
2012-01-10 23:06:57 +00:00
|
|
|
/*
|
|
|
|
* Make sure that the number of highmem pages is never larger
|
|
|
|
* than the number of the total dirtyable memory. This can only
|
|
|
|
* occur in very strange VM situations but we want to make sure
|
|
|
|
* that this does not occur.
|
|
|
|
*/
|
|
|
|
return min(x, total);
|
|
|
|
#else
|
|
|
|
return 0;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2012-01-10 23:07:44 +00:00
|
|
|
* global_dirtyable_memory - number of globally dirtyable pages
|
2012-01-10 23:06:57 +00:00
|
|
|
*
|
2012-01-10 23:07:44 +00:00
|
|
|
* Returns the global number of pages potentially available for dirty
|
|
|
|
* page cache. This is the base value for the global dirty limits.
|
2012-01-10 23:06:57 +00:00
|
|
|
*/
|
2012-04-12 20:44:20 +00:00
|
|
|
static unsigned long global_dirtyable_memory(void)
|
2012-01-10 23:06:57 +00:00
|
|
|
{
|
|
|
|
unsigned long x;
|
|
|
|
|
2014-01-29 22:05:39 +00:00
|
|
|
x = global_page_state(NR_FREE_PAGES);
|
2012-12-20 23:05:07 +00:00
|
|
|
x -= min(x, dirty_balance_reserve);
|
2012-01-10 23:06:57 +00:00
|
|
|
|
2014-01-29 22:05:41 +00:00
|
|
|
x += global_page_state(NR_INACTIVE_FILE);
|
|
|
|
x += global_page_state(NR_ACTIVE_FILE);
|
2014-01-29 22:05:39 +00:00
|
|
|
|
2012-01-10 23:06:57 +00:00
|
|
|
if (!vm_highmem_is_dirtyable)
|
|
|
|
x -= highmem_dirtyable_memory(x);
|
|
|
|
|
|
|
|
return x + 1; /* Ensure that we never return 0 */
|
|
|
|
}
|
|
|
|
|
2012-01-10 23:07:44 +00:00
|
|
|
/*
|
|
|
|
* global_dirty_limits - background-writeback and dirty-throttling thresholds
|
|
|
|
*
|
|
|
|
* Calculate the dirty thresholds based on sysctl parameters
|
|
|
|
* - vm.dirty_background_ratio or vm.dirty_background_bytes
|
|
|
|
* - vm.dirty_ratio or vm.dirty_bytes
|
|
|
|
* The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
|
|
|
|
* real-time tasks.
|
|
|
|
*/
|
|
|
|
void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
|
|
|
|
{
|
2014-08-06 23:07:31 +00:00
|
|
|
const unsigned long available_memory = global_dirtyable_memory();
|
2012-01-10 23:07:44 +00:00
|
|
|
unsigned long background;
|
|
|
|
unsigned long dirty;
|
|
|
|
struct task_struct *tsk;
|
|
|
|
|
|
|
|
if (vm_dirty_bytes)
|
|
|
|
dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
|
|
|
|
else
|
|
|
|
dirty = (vm_dirty_ratio * available_memory) / 100;
|
|
|
|
|
|
|
|
if (dirty_background_bytes)
|
|
|
|
background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
|
|
|
|
else
|
|
|
|
background = (dirty_background_ratio * available_memory) / 100;
|
|
|
|
|
|
|
|
if (background >= dirty)
|
|
|
|
background = dirty / 2;
|
|
|
|
tsk = current;
|
|
|
|
if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
|
|
|
|
background += background / 4;
|
|
|
|
dirty += dirty / 4;
|
|
|
|
}
|
|
|
|
*pbackground = background;
|
|
|
|
*pdirty = dirty;
|
|
|
|
trace_global_dirty_state(background, dirty);
|
|
|
|
}
|
|
|
|
|
mm: try to distribute dirty pages fairly across zones
The maximum number of dirty pages that exist in the system at any time is
determined by a number of pages considered dirtyable and a user-configured
percentage of those, or an absolute number in bytes.
This number of dirtyable pages is the sum of memory provided by all the
zones in the system minus their lowmem reserves and high watermarks, so
that the system can retain a healthy number of free pages without having
to reclaim dirty pages.
But there is a flaw in that we have a zoned page allocator which does not
care about the global state but rather the state of individual memory
zones. And right now there is nothing that prevents one zone from filling
up with dirty pages while other zones are spared, which frequently leads
to situations where kswapd, in order to restore the watermark of free
pages, does indeed have to write pages from that zone's LRU list. This
can interfere so badly with IO from the flusher threads that major
filesystems (btrfs, xfs, ext4) mostly ignore write requests from reclaim
already, taking away the VM's only possibility to keep such a zone
balanced, aside from hoping the flushers will soon clean pages from that
zone.
Enter per-zone dirty limits. They are to a zone's dirtyable memory what
the global limit is to the global amount of dirtyable memory, and try to
make sure that no single zone receives more than its fair share of the
globally allowed dirty pages in the first place. As the number of pages
considered dirtyable excludes the zones' lowmem reserves and high
watermarks, the maximum number of dirty pages in a zone is such that the
zone can always be balanced without requiring page cleaning.
As this is a placement decision in the page allocator and pages are
dirtied only after the allocation, this patch allows allocators to pass
__GFP_WRITE when they know in advance that the page will be written to and
become dirty soon. The page allocator will then attempt to allocate from
the first zone of the zonelist - which on NUMA is determined by the task's
NUMA memory policy - that has not exceeded its dirty limit.
At first glance, it would appear that the diversion to lower zones can
increase pressure on them, but this is not the case. With a full high
zone, allocations will be diverted to lower zones eventually, so it is
more of a shift in timing of the lower zone allocations. Workloads that
previously could fit their dirty pages completely in the higher zone may
be forced to allocate from lower zones, but the amount of pages that
"spill over" are limited themselves by the lower zones' dirty constraints,
and thus unlikely to become a problem.
For now, the problem of unfair dirty page distribution remains for NUMA
configurations where the zones allowed for allocation are in sum not big
enough to trigger the global dirty limits, wake up the flusher threads and
remedy the situation. Because of this, an allocation that could not
succeed on any of the considered zones is allowed to ignore the dirty
limits before going into direct reclaim or even failing the allocation,
until a future patch changes the global dirty throttling and flusher
thread activation so that they take individual zone states into account.
Test results
15M DMA + 3246M DMA32 + 504 Normal = 3765M memory
40% dirty ratio
16G USB thumb drive
10 runs of dd if=/dev/zero of=disk/zeroes bs=32k count=$((10 << 15))
seconds nr_vmscan_write
(stddev) min| median| max
xfs
vanilla: 549.747( 3.492) 0.000| 0.000| 0.000
patched: 550.996( 3.802) 0.000| 0.000| 0.000
fuse-ntfs
vanilla: 1183.094(53.178) 54349.000| 59341.000| 65163.000
patched: 558.049(17.914) 0.000| 0.000| 43.000
btrfs
vanilla: 573.679(14.015) 156657.000| 460178.000| 606926.000
patched: 563.365(11.368) 0.000| 0.000| 1362.000
ext4
vanilla: 561.197(15.782) 0.000|2725438.000|4143837.000
patched: 568.806(17.496) 0.000| 0.000| 0.000
Signed-off-by: Johannes Weiner <jweiner@redhat.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Michal Hocko <mhocko@suse.cz>
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shaohua Li <shaohua.li@intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Chris Mason <chris.mason@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 23:07:49 +00:00
|
|
|
/**
|
|
|
|
* zone_dirty_limit - maximum number of dirty pages allowed in a zone
|
|
|
|
* @zone: the zone
|
|
|
|
*
|
|
|
|
* Returns the maximum number of dirty pages allowed in a zone, based
|
|
|
|
* on the zone's dirtyable memory.
|
|
|
|
*/
|
|
|
|
static unsigned long zone_dirty_limit(struct zone *zone)
|
|
|
|
{
|
|
|
|
unsigned long zone_memory = zone_dirtyable_memory(zone);
|
|
|
|
struct task_struct *tsk = current;
|
|
|
|
unsigned long dirty;
|
|
|
|
|
|
|
|
if (vm_dirty_bytes)
|
|
|
|
dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
|
|
|
|
zone_memory / global_dirtyable_memory();
|
|
|
|
else
|
|
|
|
dirty = vm_dirty_ratio * zone_memory / 100;
|
|
|
|
|
|
|
|
if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
|
|
|
|
dirty += dirty / 4;
|
|
|
|
|
|
|
|
return dirty;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* zone_dirty_ok - tells whether a zone is within its dirty limits
|
|
|
|
* @zone: the zone to check
|
|
|
|
*
|
|
|
|
* Returns %true when the dirty pages in @zone are within the zone's
|
|
|
|
* dirty limit, %false if the limit is exceeded.
|
|
|
|
*/
|
|
|
|
bool zone_dirty_ok(struct zone *zone)
|
|
|
|
{
|
|
|
|
unsigned long limit = zone_dirty_limit(zone);
|
|
|
|
|
|
|
|
return zone_page_state(zone, NR_FILE_DIRTY) +
|
|
|
|
zone_page_state(zone, NR_UNSTABLE_NFS) +
|
|
|
|
zone_page_state(zone, NR_WRITEBACK) <= limit;
|
|
|
|
}
|
|
|
|
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
int dirty_background_ratio_handler(struct ctl_table *table, int write,
|
2009-09-23 22:57:19 +00:00
|
|
|
void __user *buffer, size_t *lenp,
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
loff_t *ppos)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
2009-09-23 22:57:19 +00:00
|
|
|
ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
if (ret == 0 && write)
|
|
|
|
dirty_background_bytes = 0;
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
int dirty_background_bytes_handler(struct ctl_table *table, int write,
|
2009-09-23 22:57:19 +00:00
|
|
|
void __user *buffer, size_t *lenp,
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
loff_t *ppos)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
2009-09-23 22:57:19 +00:00
|
|
|
ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
if (ret == 0 && write)
|
|
|
|
dirty_background_ratio = 0;
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2007-10-17 06:25:50 +00:00
|
|
|
int dirty_ratio_handler(struct ctl_table *table, int write,
|
2009-09-23 22:57:19 +00:00
|
|
|
void __user *buffer, size_t *lenp,
|
2007-10-17 06:25:50 +00:00
|
|
|
loff_t *ppos)
|
|
|
|
{
|
|
|
|
int old_ratio = vm_dirty_ratio;
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
int ret;
|
|
|
|
|
2009-09-23 22:57:19 +00:00
|
|
|
ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
|
2007-10-17 06:25:50 +00:00
|
|
|
if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
|
2012-05-24 16:59:11 +00:00
|
|
|
writeback_set_ratelimit();
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
vm_dirty_bytes = 0;
|
|
|
|
}
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
int dirty_bytes_handler(struct ctl_table *table, int write,
|
2009-09-23 22:57:19 +00:00
|
|
|
void __user *buffer, size_t *lenp,
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
loff_t *ppos)
|
|
|
|
{
|
2009-02-11 21:04:23 +00:00
|
|
|
unsigned long old_bytes = vm_dirty_bytes;
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
int ret;
|
|
|
|
|
2009-09-23 22:57:19 +00:00
|
|
|
ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
|
2012-05-24 16:59:11 +00:00
|
|
|
writeback_set_ratelimit();
|
mm: add dirty_background_bytes and dirty_bytes sysctls
This change introduces two new sysctls to /proc/sys/vm:
dirty_background_bytes and dirty_bytes.
dirty_background_bytes is the counterpart to dirty_background_ratio and
dirty_bytes is the counterpart to dirty_ratio.
With growing memory capacities of individual machines, it's no longer
sufficient to specify dirty thresholds as a percentage of the amount of
dirtyable memory over the entire system.
dirty_background_bytes and dirty_bytes specify quantities of memory, in
bytes, that represent the dirty limits for the entire system. If either
of these values is set, its value represents the amount of dirty memory
that is needed to commence either background or direct writeback.
When a `bytes' or `ratio' file is written, its counterpart becomes a
function of the written value. For example, if dirty_bytes is written to
be 8096, 8K of memory is required to commence direct writeback.
dirty_ratio is then functionally equivalent to 8K / the amount of
dirtyable memory:
dirtyable_memory = free pages + mapped pages + file cache
dirty_background_bytes = dirty_background_ratio * dirtyable_memory
-or-
dirty_background_ratio = dirty_background_bytes / dirtyable_memory
AND
dirty_bytes = dirty_ratio * dirtyable_memory
-or-
dirty_ratio = dirty_bytes / dirtyable_memory
Only one of dirty_background_bytes and dirty_background_ratio may be
specified at a time, and only one of dirty_bytes and dirty_ratio may be
specified. When one sysctl is written, the other appears as 0 when read.
The `bytes' files operate on a page size granularity since dirty limits
are compared with ZVC values, which are in page units.
Prior to this change, the minimum dirty_ratio was 5 as implemented by
get_dirty_limits() although /proc/sys/vm/dirty_ratio would show any user
written value between 0 and 100. This restriction is maintained, but
dirty_bytes has a lower limit of only one page.
Also prior to this change, the dirty_background_ratio could not equal or
exceed dirty_ratio. This restriction is maintained in addition to
restricting dirty_background_bytes. If either background threshold equals
or exceeds that of the dirty threshold, it is implicitly set to half the
dirty threshold.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:31 +00:00
|
|
|
vm_dirty_ratio = 0;
|
2007-10-17 06:25:50 +00:00
|
|
|
}
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2012-05-24 16:59:11 +00:00
|
|
|
static unsigned long wp_next_time(unsigned long cur_time)
|
|
|
|
{
|
|
|
|
cur_time += VM_COMPLETIONS_PERIOD_LEN;
|
|
|
|
/* 0 has a special meaning... */
|
|
|
|
if (!cur_time)
|
|
|
|
return 1;
|
|
|
|
return cur_time;
|
|
|
|
}
|
|
|
|
|
2007-10-17 06:25:50 +00:00
|
|
|
/*
|
|
|
|
* Increment the BDI's writeout completion count and the global writeout
|
|
|
|
* completion count. Called from test_clear_page_writeback().
|
|
|
|
*/
|
|
|
|
static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
|
|
|
|
{
|
2010-12-09 04:44:24 +00:00
|
|
|
__inc_bdi_stat(bdi, BDI_WRITTEN);
|
2012-05-24 16:59:11 +00:00
|
|
|
__fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
|
|
|
|
bdi->max_prop_frac);
|
|
|
|
/* First event after period switching was turned off? */
|
|
|
|
if (!unlikely(writeout_period_time)) {
|
|
|
|
/*
|
|
|
|
* We can race with other __bdi_writeout_inc calls here but
|
|
|
|
* it does not cause any harm since the resulting time when
|
|
|
|
* timer will fire and what is in writeout_period_time will be
|
|
|
|
* roughly the same.
|
|
|
|
*/
|
|
|
|
writeout_period_time = wp_next_time(jiffies);
|
|
|
|
mod_timer(&writeout_period_timer, writeout_period_time);
|
|
|
|
}
|
2007-10-17 06:25:50 +00:00
|
|
|
}
|
|
|
|
|
2008-04-30 07:54:37 +00:00
|
|
|
void bdi_writeout_inc(struct backing_dev_info *bdi)
|
|
|
|
{
|
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
local_irq_save(flags);
|
|
|
|
__bdi_writeout_inc(bdi);
|
|
|
|
local_irq_restore(flags);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(bdi_writeout_inc);
|
|
|
|
|
2007-10-17 06:25:50 +00:00
|
|
|
/*
|
|
|
|
* Obtain an accurate fraction of the BDI's portion.
|
|
|
|
*/
|
|
|
|
static void bdi_writeout_fraction(struct backing_dev_info *bdi,
|
|
|
|
long *numerator, long *denominator)
|
|
|
|
{
|
2012-05-24 16:59:11 +00:00
|
|
|
fprop_fraction_percpu(&writeout_completions, &bdi->completions,
|
2007-10-17 06:25:50 +00:00
|
|
|
numerator, denominator);
|
|
|
|
}
|
|
|
|
|
2012-05-24 16:59:11 +00:00
|
|
|
/*
|
|
|
|
* On idle system, we can be called long after we scheduled because we use
|
|
|
|
* deferred timers so count with missed periods.
|
|
|
|
*/
|
|
|
|
static void writeout_period(unsigned long t)
|
|
|
|
{
|
|
|
|
int miss_periods = (jiffies - writeout_period_time) /
|
|
|
|
VM_COMPLETIONS_PERIOD_LEN;
|
|
|
|
|
|
|
|
if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
|
|
|
|
writeout_period_time = wp_next_time(writeout_period_time +
|
|
|
|
miss_periods * VM_COMPLETIONS_PERIOD_LEN);
|
|
|
|
mod_timer(&writeout_period_timer, writeout_period_time);
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* Aging has zeroed all fractions. Stop wasting CPU on period
|
|
|
|
* updates.
|
|
|
|
*/
|
|
|
|
writeout_period_time = 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-04-30 07:54:35 +00:00
|
|
|
/*
|
2011-11-01 00:07:05 +00:00
|
|
|
* bdi_min_ratio keeps the sum of the minimum dirty shares of all
|
|
|
|
* registered backing devices, which, for obvious reasons, can not
|
|
|
|
* exceed 100%.
|
2008-04-30 07:54:35 +00:00
|
|
|
*/
|
|
|
|
static unsigned int bdi_min_ratio;
|
|
|
|
|
|
|
|
int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
|
2009-09-14 11:12:40 +00:00
|
|
|
spin_lock_bh(&bdi_lock);
|
2008-04-30 07:54:36 +00:00
|
|
|
if (min_ratio > bdi->max_ratio) {
|
2008-04-30 07:54:35 +00:00
|
|
|
ret = -EINVAL;
|
2008-04-30 07:54:36 +00:00
|
|
|
} else {
|
|
|
|
min_ratio -= bdi->min_ratio;
|
|
|
|
if (bdi_min_ratio + min_ratio < 100) {
|
|
|
|
bdi_min_ratio += min_ratio;
|
|
|
|
bdi->min_ratio += min_ratio;
|
|
|
|
} else {
|
|
|
|
ret = -EINVAL;
|
|
|
|
}
|
|
|
|
}
|
2009-09-14 11:12:40 +00:00
|
|
|
spin_unlock_bh(&bdi_lock);
|
2008-04-30 07:54:36 +00:00
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
|
|
|
|
if (max_ratio > 100)
|
|
|
|
return -EINVAL;
|
|
|
|
|
2009-09-14 11:12:40 +00:00
|
|
|
spin_lock_bh(&bdi_lock);
|
2008-04-30 07:54:36 +00:00
|
|
|
if (bdi->min_ratio > max_ratio) {
|
|
|
|
ret = -EINVAL;
|
|
|
|
} else {
|
|
|
|
bdi->max_ratio = max_ratio;
|
2012-05-24 16:59:11 +00:00
|
|
|
bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
|
2008-04-30 07:54:36 +00:00
|
|
|
}
|
2009-09-14 11:12:40 +00:00
|
|
|
spin_unlock_bh(&bdi_lock);
|
2008-04-30 07:54:35 +00:00
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
2008-04-30 07:54:36 +00:00
|
|
|
EXPORT_SYMBOL(bdi_set_max_ratio);
|
2008-04-30 07:54:35 +00:00
|
|
|
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
static unsigned long dirty_freerun_ceiling(unsigned long thresh,
|
|
|
|
unsigned long bg_thresh)
|
|
|
|
{
|
|
|
|
return (thresh + bg_thresh) / 2;
|
|
|
|
}
|
|
|
|
|
2011-06-20 04:18:42 +00:00
|
|
|
static unsigned long hard_dirty_limit(unsigned long thresh)
|
|
|
|
{
|
|
|
|
return max(thresh, global_dirty_limit);
|
|
|
|
}
|
|
|
|
|
2011-03-02 23:14:34 +00:00
|
|
|
/**
|
2010-08-11 21:17:40 +00:00
|
|
|
* bdi_dirty_limit - @bdi's share of dirty throttling threshold
|
2011-03-02 23:14:34 +00:00
|
|
|
* @bdi: the backing_dev_info to query
|
|
|
|
* @dirty: global dirty limit in pages
|
2010-08-11 21:17:40 +00:00
|
|
|
*
|
2011-03-02 23:14:34 +00:00
|
|
|
* Returns @bdi's dirty limit in pages. The term "dirty" in the context of
|
|
|
|
* dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
|
2011-11-23 17:44:41 +00:00
|
|
|
*
|
|
|
|
* Note that balance_dirty_pages() will only seriously take it as a hard limit
|
|
|
|
* when sleeping max_pause per page is not enough to keep the dirty pages under
|
|
|
|
* control. For example, when the device is completely stalled due to some error
|
|
|
|
* conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
|
|
|
|
* In the other normal situations, it acts more gently by throttling the tasks
|
|
|
|
* more (rather than completely block them) when the bdi dirty pages go high.
|
2010-08-11 21:17:40 +00:00
|
|
|
*
|
2011-03-02 23:14:34 +00:00
|
|
|
* It allocates high/low dirty limits to fast/slow devices, in order to prevent
|
2010-08-11 21:17:40 +00:00
|
|
|
* - starving fast devices
|
|
|
|
* - piling up dirty pages (that will take long time to sync) on slow devices
|
|
|
|
*
|
|
|
|
* The bdi's share of dirty limit will be adapting to its throughput and
|
|
|
|
* bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
|
|
|
|
*/
|
|
|
|
unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
|
2010-08-11 21:17:39 +00:00
|
|
|
{
|
|
|
|
u64 bdi_dirty;
|
|
|
|
long numerator, denominator;
|
2007-10-17 06:25:50 +00:00
|
|
|
|
2010-08-11 21:17:39 +00:00
|
|
|
/*
|
|
|
|
* Calculate this BDI's share of the dirty ratio.
|
|
|
|
*/
|
|
|
|
bdi_writeout_fraction(bdi, &numerator, &denominator);
|
2007-10-17 06:25:50 +00:00
|
|
|
|
2010-08-11 21:17:39 +00:00
|
|
|
bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
|
|
|
|
bdi_dirty *= numerator;
|
|
|
|
do_div(bdi_dirty, denominator);
|
2007-10-17 06:25:50 +00:00
|
|
|
|
2010-08-11 21:17:39 +00:00
|
|
|
bdi_dirty += (dirty * bdi->min_ratio) / 100;
|
|
|
|
if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
|
|
|
|
bdi_dirty = dirty * bdi->max_ratio / 100;
|
|
|
|
|
|
|
|
return bdi_dirty;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
/*
|
|
|
|
* setpoint - dirty 3
|
|
|
|
* f(dirty) := 1.0 + (----------------)
|
|
|
|
* limit - setpoint
|
|
|
|
*
|
|
|
|
* it's a 3rd order polynomial that subjects to
|
|
|
|
*
|
|
|
|
* (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
|
|
|
|
* (2) f(setpoint) = 1.0 => the balance point
|
|
|
|
* (3) f(limit) = 0 => the hard limit
|
|
|
|
* (4) df/dx <= 0 => negative feedback control
|
|
|
|
* (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
|
|
|
|
* => fast response on large errors; small oscillation near setpoint
|
|
|
|
*/
|
2014-05-06 19:50:01 +00:00
|
|
|
static long long pos_ratio_polynom(unsigned long setpoint,
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
unsigned long dirty,
|
|
|
|
unsigned long limit)
|
|
|
|
{
|
|
|
|
long long pos_ratio;
|
|
|
|
long x;
|
|
|
|
|
2014-05-06 19:50:01 +00:00
|
|
|
x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
limit - setpoint + 1);
|
|
|
|
pos_ratio = x;
|
|
|
|
pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
|
|
|
|
pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
|
|
|
|
pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
|
|
|
|
|
|
|
|
return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
|
|
|
|
}
|
|
|
|
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
/*
|
|
|
|
* Dirty position control.
|
|
|
|
*
|
|
|
|
* (o) global/bdi setpoints
|
|
|
|
*
|
|
|
|
* We want the dirty pages be balanced around the global/bdi setpoints.
|
|
|
|
* When the number of dirty pages is higher/lower than the setpoint, the
|
|
|
|
* dirty position control ratio (and hence task dirty ratelimit) will be
|
|
|
|
* decreased/increased to bring the dirty pages back to the setpoint.
|
|
|
|
*
|
|
|
|
* pos_ratio = 1 << RATELIMIT_CALC_SHIFT
|
|
|
|
*
|
|
|
|
* if (dirty < setpoint) scale up pos_ratio
|
|
|
|
* if (dirty > setpoint) scale down pos_ratio
|
|
|
|
*
|
|
|
|
* if (bdi_dirty < bdi_setpoint) scale up pos_ratio
|
|
|
|
* if (bdi_dirty > bdi_setpoint) scale down pos_ratio
|
|
|
|
*
|
|
|
|
* task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
|
|
|
|
*
|
|
|
|
* (o) global control line
|
|
|
|
*
|
|
|
|
* ^ pos_ratio
|
|
|
|
* |
|
|
|
|
* | |<===== global dirty control scope ======>|
|
|
|
|
* 2.0 .............*
|
|
|
|
* | .*
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* 1.0 ................................*
|
|
|
|
* | . . *
|
|
|
|
* | . . *
|
|
|
|
* | . . *
|
|
|
|
* | . . *
|
|
|
|
* | . . *
|
|
|
|
* 0 +------------.------------------.----------------------*------------->
|
|
|
|
* freerun^ setpoint^ limit^ dirty pages
|
|
|
|
*
|
|
|
|
* (o) bdi control line
|
|
|
|
*
|
|
|
|
* ^ pos_ratio
|
|
|
|
* |
|
|
|
|
* | *
|
|
|
|
* | *
|
|
|
|
* | *
|
|
|
|
* | *
|
|
|
|
* | * |<=========== span ============>|
|
|
|
|
* 1.0 .......................*
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* | . *
|
|
|
|
* 1/4 ...............................................* * * * * * * * * * * *
|
|
|
|
* | . .
|
|
|
|
* | . .
|
|
|
|
* | . .
|
|
|
|
* 0 +----------------------.-------------------------------.------------->
|
|
|
|
* bdi_setpoint^ x_intercept^
|
|
|
|
*
|
|
|
|
* The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
|
|
|
|
* be smoothly throttled down to normal if it starts high in situations like
|
|
|
|
* - start writing to a slow SD card and a fast disk at the same time. The SD
|
|
|
|
* card's bdi_dirty may rush to many times higher than bdi_setpoint.
|
|
|
|
* - the bdi dirty thresh drops quickly due to change of JBOD workload
|
|
|
|
*/
|
|
|
|
static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
|
|
|
|
unsigned long thresh,
|
|
|
|
unsigned long bg_thresh,
|
|
|
|
unsigned long dirty,
|
|
|
|
unsigned long bdi_thresh,
|
|
|
|
unsigned long bdi_dirty)
|
|
|
|
{
|
|
|
|
unsigned long write_bw = bdi->avg_write_bandwidth;
|
|
|
|
unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
|
|
|
|
unsigned long limit = hard_dirty_limit(thresh);
|
|
|
|
unsigned long x_intercept;
|
|
|
|
unsigned long setpoint; /* dirty pages' target balance point */
|
|
|
|
unsigned long bdi_setpoint;
|
|
|
|
unsigned long span;
|
|
|
|
long long pos_ratio; /* for scaling up/down the rate limit */
|
|
|
|
long x;
|
|
|
|
|
|
|
|
if (unlikely(dirty >= limit))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* global setpoint
|
|
|
|
*
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
* See comment for pos_ratio_polynom().
|
|
|
|
*/
|
|
|
|
setpoint = (freerun + limit) / 2;
|
|
|
|
pos_ratio = pos_ratio_polynom(setpoint, dirty, limit);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The strictlimit feature is a tool preventing mistrusted filesystems
|
|
|
|
* from growing a large number of dirty pages before throttling. For
|
|
|
|
* such filesystems balance_dirty_pages always checks bdi counters
|
|
|
|
* against bdi limits. Even if global "nr_dirty" is under "freerun".
|
|
|
|
* This is especially important for fuse which sets bdi->max_ratio to
|
|
|
|
* 1% by default. Without strictlimit feature, fuse writeback may
|
|
|
|
* consume arbitrary amount of RAM because it is accounted in
|
|
|
|
* NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
*
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
* Here, in bdi_position_ratio(), we calculate pos_ratio based on
|
|
|
|
* two values: bdi_dirty and bdi_thresh. Let's consider an example:
|
|
|
|
* total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
|
|
|
|
* limits are set by default to 10% and 20% (background and throttle).
|
|
|
|
* Then bdi_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
|
|
|
|
* bdi_dirty_limit(bdi, bg_thresh) is about ~4K pages. bdi_setpoint is
|
|
|
|
* about ~6K pages (as the average of background and throttle bdi
|
|
|
|
* limits). The 3rd order polynomial will provide positive feedback if
|
|
|
|
* bdi_dirty is under bdi_setpoint and vice versa.
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
*
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
* Note, that we cannot use global counters in these calculations
|
|
|
|
* because we want to throttle process writing to a strictlimit BDI
|
|
|
|
* much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
|
|
|
|
* in the example above).
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
*/
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
|
|
|
|
long long bdi_pos_ratio;
|
|
|
|
unsigned long bdi_bg_thresh;
|
|
|
|
|
|
|
|
if (bdi_dirty < 8)
|
|
|
|
return min_t(long long, pos_ratio * 2,
|
|
|
|
2 << RATELIMIT_CALC_SHIFT);
|
|
|
|
|
|
|
|
if (bdi_dirty >= bdi_thresh)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
bdi_bg_thresh = div_u64((u64)bdi_thresh * bg_thresh, thresh);
|
|
|
|
bdi_setpoint = dirty_freerun_ceiling(bdi_thresh,
|
|
|
|
bdi_bg_thresh);
|
|
|
|
|
|
|
|
if (bdi_setpoint == 0 || bdi_setpoint == bdi_thresh)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
bdi_pos_ratio = pos_ratio_polynom(bdi_setpoint, bdi_dirty,
|
|
|
|
bdi_thresh);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Typically, for strictlimit case, bdi_setpoint << setpoint
|
|
|
|
* and pos_ratio >> bdi_pos_ratio. In the other words global
|
|
|
|
* state ("dirty") is not limiting factor and we have to
|
|
|
|
* make decision based on bdi counters. But there is an
|
|
|
|
* important case when global pos_ratio should get precedence:
|
|
|
|
* global limits are exceeded (e.g. due to activities on other
|
|
|
|
* BDIs) while given strictlimit BDI is below limit.
|
|
|
|
*
|
|
|
|
* "pos_ratio * bdi_pos_ratio" would work for the case above,
|
|
|
|
* but it would look too non-natural for the case of all
|
|
|
|
* activity in the system coming from a single strictlimit BDI
|
|
|
|
* with bdi->max_ratio == 100%.
|
|
|
|
*
|
|
|
|
* Note that min() below somewhat changes the dynamics of the
|
|
|
|
* control system. Normally, pos_ratio value can be well over 3
|
|
|
|
* (when globally we are at freerun and bdi is well below bdi
|
|
|
|
* setpoint). Now the maximum pos_ratio in the same situation
|
|
|
|
* is 2. We might want to tweak this if we observe the control
|
|
|
|
* system is too slow to adapt.
|
|
|
|
*/
|
|
|
|
return min(pos_ratio, bdi_pos_ratio);
|
|
|
|
}
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* We have computed basic pos_ratio above based on global situation. If
|
|
|
|
* the bdi is over/under its share of dirty pages, we want to scale
|
|
|
|
* pos_ratio further down/up. That is done by the following mechanism.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* bdi setpoint
|
|
|
|
*
|
|
|
|
* f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
|
|
|
|
*
|
|
|
|
* x_intercept - bdi_dirty
|
|
|
|
* := --------------------------
|
|
|
|
* x_intercept - bdi_setpoint
|
|
|
|
*
|
|
|
|
* The main bdi control line is a linear function that subjects to
|
|
|
|
*
|
|
|
|
* (1) f(bdi_setpoint) = 1.0
|
|
|
|
* (2) k = - 1 / (8 * write_bw) (in single bdi case)
|
|
|
|
* or equally: x_intercept = bdi_setpoint + 8 * write_bw
|
|
|
|
*
|
|
|
|
* For single bdi case, the dirty pages are observed to fluctuate
|
|
|
|
* regularly within range
|
|
|
|
* [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
|
|
|
|
* for various filesystems, where (2) can yield in a reasonable 12.5%
|
|
|
|
* fluctuation range for pos_ratio.
|
|
|
|
*
|
|
|
|
* For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
|
|
|
|
* own size, so move the slope over accordingly and choose a slope that
|
|
|
|
* yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
|
|
|
|
*/
|
|
|
|
if (unlikely(bdi_thresh > thresh))
|
|
|
|
bdi_thresh = thresh;
|
2011-11-23 17:44:41 +00:00
|
|
|
/*
|
|
|
|
* It's very possible that bdi_thresh is close to 0 not because the
|
|
|
|
* device is slow, but that it has remained inactive for long time.
|
|
|
|
* Honour such devices a reasonable good (hopefully IO efficient)
|
|
|
|
* threshold, so that the occasional writes won't be blocked and active
|
|
|
|
* writes can rampup the threshold quickly.
|
|
|
|
*/
|
2011-08-05 04:16:46 +00:00
|
|
|
bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
/*
|
|
|
|
* scale global setpoint to bdi's:
|
|
|
|
* bdi_setpoint = setpoint * bdi_thresh / thresh
|
|
|
|
*/
|
|
|
|
x = div_u64((u64)bdi_thresh << 16, thresh + 1);
|
|
|
|
bdi_setpoint = setpoint * (u64)x >> 16;
|
|
|
|
/*
|
|
|
|
* Use span=(8*write_bw) in single bdi case as indicated by
|
|
|
|
* (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
|
|
|
|
*
|
|
|
|
* bdi_thresh thresh - bdi_thresh
|
|
|
|
* span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
|
|
|
|
* thresh thresh
|
|
|
|
*/
|
|
|
|
span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
|
|
|
|
x_intercept = bdi_setpoint + span;
|
|
|
|
|
|
|
|
if (bdi_dirty < x_intercept - span / 4) {
|
2014-05-06 19:50:01 +00:00
|
|
|
pos_ratio = div64_u64(pos_ratio * (x_intercept - bdi_dirty),
|
2011-10-11 23:06:33 +00:00
|
|
|
x_intercept - bdi_setpoint + 1);
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
} else
|
|
|
|
pos_ratio /= 4;
|
|
|
|
|
2011-08-05 04:16:46 +00:00
|
|
|
/*
|
|
|
|
* bdi reserve area, safeguard against dirty pool underrun and disk idle
|
|
|
|
* It may push the desired control point of global dirty pages higher
|
|
|
|
* than setpoint.
|
|
|
|
*/
|
|
|
|
x_intercept = bdi_thresh / 2;
|
|
|
|
if (bdi_dirty < x_intercept) {
|
2011-10-11 23:06:33 +00:00
|
|
|
if (bdi_dirty > x_intercept / 8)
|
|
|
|
pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
|
|
|
|
else
|
2011-08-05 04:16:46 +00:00
|
|
|
pos_ratio *= 8;
|
|
|
|
}
|
|
|
|
|
writeback: dirty position control
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>
2011-03-02 22:04:18 +00:00
|
|
|
return pos_ratio;
|
|
|
|
}
|
|
|
|
|
2010-08-29 17:22:30 +00:00
|
|
|
static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
|
|
|
|
unsigned long elapsed,
|
|
|
|
unsigned long written)
|
|
|
|
{
|
|
|
|
const unsigned long period = roundup_pow_of_two(3 * HZ);
|
|
|
|
unsigned long avg = bdi->avg_write_bandwidth;
|
|
|
|
unsigned long old = bdi->write_bandwidth;
|
|
|
|
u64 bw;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* bw = written * HZ / elapsed
|
|
|
|
*
|
|
|
|
* bw * elapsed + write_bandwidth * (period - elapsed)
|
|
|
|
* write_bandwidth = ---------------------------------------------------
|
|
|
|
* period
|
|
|
|
*/
|
|
|
|
bw = written - bdi->written_stamp;
|
|
|
|
bw *= HZ;
|
|
|
|
if (unlikely(elapsed > period)) {
|
|
|
|
do_div(bw, elapsed);
|
|
|
|
avg = bw;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
bw += (u64)bdi->write_bandwidth * (period - elapsed);
|
|
|
|
bw >>= ilog2(period);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* one more level of smoothing, for filtering out sudden spikes
|
|
|
|
*/
|
|
|
|
if (avg > old && old >= (unsigned long)bw)
|
|
|
|
avg -= (avg - old) >> 3;
|
|
|
|
|
|
|
|
if (avg < old && old <= (unsigned long)bw)
|
|
|
|
avg += (old - avg) >> 3;
|
|
|
|
|
|
|
|
out:
|
|
|
|
bdi->write_bandwidth = bw;
|
|
|
|
bdi->avg_write_bandwidth = avg;
|
|
|
|
}
|
|
|
|
|
2011-03-02 21:54:09 +00:00
|
|
|
/*
|
|
|
|
* The global dirtyable memory and dirty threshold could be suddenly knocked
|
|
|
|
* down by a large amount (eg. on the startup of KVM in a swapless system).
|
|
|
|
* This may throw the system into deep dirty exceeded state and throttle
|
|
|
|
* heavy/light dirtiers alike. To retain good responsiveness, maintain
|
|
|
|
* global_dirty_limit for tracking slowly down to the knocked down dirty
|
|
|
|
* threshold.
|
|
|
|
*/
|
|
|
|
static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
|
|
|
|
{
|
|
|
|
unsigned long limit = global_dirty_limit;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Follow up in one step.
|
|
|
|
*/
|
|
|
|
if (limit < thresh) {
|
|
|
|
limit = thresh;
|
|
|
|
goto update;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Follow down slowly. Use the higher one as the target, because thresh
|
|
|
|
* may drop below dirty. This is exactly the reason to introduce
|
|
|
|
* global_dirty_limit which is guaranteed to lie above the dirty pages.
|
|
|
|
*/
|
|
|
|
thresh = max(thresh, dirty);
|
|
|
|
if (limit > thresh) {
|
|
|
|
limit -= (limit - thresh) >> 5;
|
|
|
|
goto update;
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
update:
|
|
|
|
global_dirty_limit = limit;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void global_update_bandwidth(unsigned long thresh,
|
|
|
|
unsigned long dirty,
|
|
|
|
unsigned long now)
|
|
|
|
{
|
|
|
|
static DEFINE_SPINLOCK(dirty_lock);
|
|
|
|
static unsigned long update_time;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* check locklessly first to optimize away locking for the most time
|
|
|
|
*/
|
|
|
|
if (time_before(now, update_time + BANDWIDTH_INTERVAL))
|
|
|
|
return;
|
|
|
|
|
|
|
|
spin_lock(&dirty_lock);
|
|
|
|
if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
|
|
|
|
update_dirty_limit(thresh, dirty);
|
|
|
|
update_time = now;
|
|
|
|
}
|
|
|
|
spin_unlock(&dirty_lock);
|
|
|
|
}
|
|
|
|
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
/*
|
|
|
|
* Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
|
|
|
|
*
|
|
|
|
* Normal bdi tasks will be curbed at or below it in long term.
|
|
|
|
* Obviously it should be around (write_bw / N) when there are N dd tasks.
|
|
|
|
*/
|
|
|
|
static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
|
|
|
|
unsigned long thresh,
|
|
|
|
unsigned long bg_thresh,
|
|
|
|
unsigned long dirty,
|
|
|
|
unsigned long bdi_thresh,
|
|
|
|
unsigned long bdi_dirty,
|
|
|
|
unsigned long dirtied,
|
|
|
|
unsigned long elapsed)
|
|
|
|
{
|
writeback: stabilize bdi->dirty_ratelimit
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>
2011-08-26 21:53:24 +00:00
|
|
|
unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
|
|
|
|
unsigned long limit = hard_dirty_limit(thresh);
|
|
|
|
unsigned long setpoint = (freerun + limit) / 2;
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
unsigned long write_bw = bdi->avg_write_bandwidth;
|
|
|
|
unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
|
|
|
|
unsigned long dirty_rate;
|
|
|
|
unsigned long task_ratelimit;
|
|
|
|
unsigned long balanced_dirty_ratelimit;
|
|
|
|
unsigned long pos_ratio;
|
writeback: stabilize bdi->dirty_ratelimit
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>
2011-08-26 21:53:24 +00:00
|
|
|
unsigned long step;
|
|
|
|
unsigned long x;
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* The dirty rate will match the writeout rate in long term, except
|
|
|
|
* when dirty pages are truncated by userspace or re-dirtied by FS.
|
|
|
|
*/
|
|
|
|
dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
|
|
|
|
|
|
|
|
pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
|
|
|
|
bdi_thresh, bdi_dirty);
|
|
|
|
/*
|
|
|
|
* task_ratelimit reflects each dd's dirty rate for the past 200ms.
|
|
|
|
*/
|
|
|
|
task_ratelimit = (u64)dirty_ratelimit *
|
|
|
|
pos_ratio >> RATELIMIT_CALC_SHIFT;
|
|
|
|
task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* A linear estimation of the "balanced" throttle rate. The theory is,
|
|
|
|
* if there are N dd tasks, each throttled at task_ratelimit, the bdi's
|
|
|
|
* dirty_rate will be measured to be (N * task_ratelimit). So the below
|
|
|
|
* formula will yield the balanced rate limit (write_bw / N).
|
|
|
|
*
|
|
|
|
* Note that the expanded form is not a pure rate feedback:
|
|
|
|
* rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
|
|
|
|
* but also takes pos_ratio into account:
|
|
|
|
* rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
|
|
|
|
*
|
|
|
|
* (1) is not realistic because pos_ratio also takes part in balancing
|
|
|
|
* the dirty rate. Consider the state
|
|
|
|
* pos_ratio = 0.5 (3)
|
|
|
|
* rate = 2 * (write_bw / N) (4)
|
|
|
|
* If (1) is used, it will stuck in that state! Because each dd will
|
|
|
|
* be throttled at
|
|
|
|
* task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
|
|
|
|
* yielding
|
|
|
|
* dirty_rate = N * task_ratelimit = write_bw (6)
|
|
|
|
* put (6) into (1) we get
|
|
|
|
* rate_(i+1) = rate_(i) (7)
|
|
|
|
*
|
|
|
|
* So we end up using (2) to always keep
|
|
|
|
* rate_(i+1) ~= (write_bw / N) (8)
|
|
|
|
* regardless of the value of pos_ratio. As long as (8) is satisfied,
|
|
|
|
* pos_ratio is able to drive itself to 1.0, which is not only where
|
|
|
|
* the dirty count meet the setpoint, but also where the slope of
|
|
|
|
* pos_ratio is most flat and hence task_ratelimit is least fluctuated.
|
|
|
|
*/
|
|
|
|
balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
|
|
|
|
dirty_rate | 1);
|
2011-08-03 20:30:36 +00:00
|
|
|
/*
|
|
|
|
* balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
|
|
|
|
*/
|
|
|
|
if (unlikely(balanced_dirty_ratelimit > write_bw))
|
|
|
|
balanced_dirty_ratelimit = write_bw;
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
|
writeback: stabilize bdi->dirty_ratelimit
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>
2011-08-26 21:53:24 +00:00
|
|
|
/*
|
|
|
|
* We could safely do this and return immediately:
|
|
|
|
*
|
|
|
|
* bdi->dirty_ratelimit = balanced_dirty_ratelimit;
|
|
|
|
*
|
|
|
|
* However to get a more stable dirty_ratelimit, the below elaborated
|
2012-06-09 03:10:55 +00:00
|
|
|
* code makes use of task_ratelimit to filter out singular points and
|
writeback: stabilize bdi->dirty_ratelimit
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>
2011-08-26 21:53:24 +00:00
|
|
|
* limit the step size.
|
|
|
|
*
|
|
|
|
* The below code essentially only uses the relative value of
|
|
|
|
*
|
|
|
|
* task_ratelimit - dirty_ratelimit
|
|
|
|
* = (pos_ratio - 1) * dirty_ratelimit
|
|
|
|
*
|
|
|
|
* which reflects the direction and size of dirty position error.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* dirty_ratelimit will follow balanced_dirty_ratelimit iff
|
|
|
|
* task_ratelimit is on the same side of dirty_ratelimit, too.
|
|
|
|
* For example, when
|
|
|
|
* - dirty_ratelimit > balanced_dirty_ratelimit
|
|
|
|
* - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
|
|
|
|
* lowering dirty_ratelimit will help meet both the position and rate
|
|
|
|
* control targets. Otherwise, don't update dirty_ratelimit if it will
|
|
|
|
* only help meet the rate target. After all, what the users ultimately
|
|
|
|
* feel and care are stable dirty rate and small position error.
|
|
|
|
*
|
|
|
|
* |task_ratelimit - dirty_ratelimit| is used to limit the step size
|
2012-06-09 03:10:55 +00:00
|
|
|
* and filter out the singular points of balanced_dirty_ratelimit. Which
|
writeback: stabilize bdi->dirty_ratelimit
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>
2011-08-26 21:53:24 +00:00
|
|
|
* keeps jumping around randomly and can even leap far away at times
|
|
|
|
* due to the small 200ms estimation period of dirty_rate (we want to
|
|
|
|
* keep that period small to reduce time lags).
|
|
|
|
*/
|
|
|
|
step = 0;
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* For strictlimit case, calculations above were based on bdi counters
|
|
|
|
* and limits (starting from pos_ratio = bdi_position_ratio() and up to
|
|
|
|
* balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
|
|
|
|
* Hence, to calculate "step" properly, we have to use bdi_dirty as
|
|
|
|
* "dirty" and bdi_setpoint as "setpoint".
|
|
|
|
*
|
|
|
|
* We rampup dirty_ratelimit forcibly if bdi_dirty is low because
|
|
|
|
* it's possible that bdi_thresh is close to zero due to inactivity
|
|
|
|
* of backing device (see the implementation of bdi_dirty_limit()).
|
|
|
|
*/
|
|
|
|
if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
|
|
|
|
dirty = bdi_dirty;
|
|
|
|
if (bdi_dirty < 8)
|
|
|
|
setpoint = bdi_dirty + 1;
|
|
|
|
else
|
|
|
|
setpoint = (bdi_thresh +
|
|
|
|
bdi_dirty_limit(bdi, bg_thresh)) / 2;
|
|
|
|
}
|
|
|
|
|
writeback: stabilize bdi->dirty_ratelimit
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>
2011-08-26 21:53:24 +00:00
|
|
|
if (dirty < setpoint) {
|
2014-10-09 22:28:15 +00:00
|
|
|
x = min3(bdi->balanced_dirty_ratelimit,
|
|
|
|
balanced_dirty_ratelimit, task_ratelimit);
|
writeback: stabilize bdi->dirty_ratelimit
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>
2011-08-26 21:53:24 +00:00
|
|
|
if (dirty_ratelimit < x)
|
|
|
|
step = x - dirty_ratelimit;
|
|
|
|
} else {
|
2014-10-09 22:28:15 +00:00
|
|
|
x = max3(bdi->balanced_dirty_ratelimit,
|
|
|
|
balanced_dirty_ratelimit, task_ratelimit);
|
writeback: stabilize bdi->dirty_ratelimit
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>
2011-08-26 21:53:24 +00:00
|
|
|
if (dirty_ratelimit > x)
|
|
|
|
step = dirty_ratelimit - x;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Don't pursue 100% rate matching. It's impossible since the balanced
|
|
|
|
* rate itself is constantly fluctuating. So decrease the track speed
|
|
|
|
* when it gets close to the target. Helps eliminate pointless tremors.
|
|
|
|
*/
|
|
|
|
step >>= dirty_ratelimit / (2 * step + 1);
|
|
|
|
/*
|
|
|
|
* Limit the tracking speed to avoid overshooting.
|
|
|
|
*/
|
|
|
|
step = (step + 7) / 8;
|
|
|
|
|
|
|
|
if (dirty_ratelimit < balanced_dirty_ratelimit)
|
|
|
|
dirty_ratelimit += step;
|
|
|
|
else
|
|
|
|
dirty_ratelimit -= step;
|
|
|
|
|
|
|
|
bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
|
|
|
|
bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
|
2011-03-02 23:22:49 +00:00
|
|
|
|
|
|
|
trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
}
|
|
|
|
|
2010-08-29 17:22:30 +00:00
|
|
|
void __bdi_update_bandwidth(struct backing_dev_info *bdi,
|
2011-03-02 21:54:09 +00:00
|
|
|
unsigned long thresh,
|
2011-10-04 02:46:17 +00:00
|
|
|
unsigned long bg_thresh,
|
2011-03-02 21:54:09 +00:00
|
|
|
unsigned long dirty,
|
|
|
|
unsigned long bdi_thresh,
|
|
|
|
unsigned long bdi_dirty,
|
2010-08-29 17:22:30 +00:00
|
|
|
unsigned long start_time)
|
|
|
|
{
|
|
|
|
unsigned long now = jiffies;
|
|
|
|
unsigned long elapsed = now - bdi->bw_time_stamp;
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
unsigned long dirtied;
|
2010-08-29 17:22:30 +00:00
|
|
|
unsigned long written;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* rate-limit, only update once every 200ms.
|
|
|
|
*/
|
|
|
|
if (elapsed < BANDWIDTH_INTERVAL)
|
|
|
|
return;
|
|
|
|
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
|
2010-08-29 17:22:30 +00:00
|
|
|
written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Skip quiet periods when disk bandwidth is under-utilized.
|
|
|
|
* (at least 1s idle time between two flusher runs)
|
|
|
|
*/
|
|
|
|
if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
|
|
|
|
goto snapshot;
|
|
|
|
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
if (thresh) {
|
2011-03-02 21:54:09 +00:00
|
|
|
global_update_bandwidth(thresh, dirty, now);
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
|
|
|
|
bdi_thresh, bdi_dirty,
|
|
|
|
dirtied, elapsed);
|
|
|
|
}
|
2010-08-29 17:22:30 +00:00
|
|
|
bdi_update_write_bandwidth(bdi, elapsed, written);
|
|
|
|
|
|
|
|
snapshot:
|
writeback: dirty rate control
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>
2011-06-12 16:51:31 +00:00
|
|
|
bdi->dirtied_stamp = dirtied;
|
2010-08-29 17:22:30 +00:00
|
|
|
bdi->written_stamp = written;
|
|
|
|
bdi->bw_time_stamp = now;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void bdi_update_bandwidth(struct backing_dev_info *bdi,
|
2011-03-02 21:54:09 +00:00
|
|
|
unsigned long thresh,
|
2011-10-04 02:46:17 +00:00
|
|
|
unsigned long bg_thresh,
|
2011-03-02 21:54:09 +00:00
|
|
|
unsigned long dirty,
|
|
|
|
unsigned long bdi_thresh,
|
|
|
|
unsigned long bdi_dirty,
|
2010-08-29 17:22:30 +00:00
|
|
|
unsigned long start_time)
|
|
|
|
{
|
|
|
|
if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
|
|
|
|
return;
|
|
|
|
spin_lock(&bdi->wb.list_lock);
|
2011-10-04 02:46:17 +00:00
|
|
|
__bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
|
|
|
|
bdi_thresh, bdi_dirty, start_time);
|
2010-08-29 17:22:30 +00:00
|
|
|
spin_unlock(&bdi->wb.list_lock);
|
|
|
|
}
|
|
|
|
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
/*
|
2012-12-12 00:00:21 +00:00
|
|
|
* After a task dirtied this many pages, balance_dirty_pages_ratelimited()
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
* will look to see if it needs to start dirty throttling.
|
|
|
|
*
|
|
|
|
* If dirty_poll_interval is too low, big NUMA machines will call the expensive
|
|
|
|
* global_page_state() too often. So scale it near-sqrt to the safety margin
|
|
|
|
* (the number of pages we may dirty without exceeding the dirty limits).
|
|
|
|
*/
|
|
|
|
static unsigned long dirty_poll_interval(unsigned long dirty,
|
|
|
|
unsigned long thresh)
|
|
|
|
{
|
|
|
|
if (thresh > dirty)
|
|
|
|
return 1UL << (ilog2(thresh - dirty) >> 1);
|
|
|
|
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
2013-10-16 20:47:03 +00:00
|
|
|
static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
|
|
|
|
unsigned long bdi_dirty)
|
2011-06-12 01:21:43 +00:00
|
|
|
{
|
2013-10-16 20:47:03 +00:00
|
|
|
unsigned long bw = bdi->avg_write_bandwidth;
|
|
|
|
unsigned long t;
|
2011-06-12 01:21:43 +00:00
|
|
|
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
/*
|
|
|
|
* Limit pause time for small memory systems. If sleeping for too long
|
|
|
|
* time, a small pool of dirty/writeback pages may go empty and disk go
|
|
|
|
* idle.
|
|
|
|
*
|
|
|
|
* 8 serves as the safety ratio.
|
|
|
|
*/
|
|
|
|
t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
|
|
|
|
t++;
|
|
|
|
|
2013-10-16 20:47:03 +00:00
|
|
|
return min_t(unsigned long, t, MAX_PAUSE);
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static long bdi_min_pause(struct backing_dev_info *bdi,
|
|
|
|
long max_pause,
|
|
|
|
unsigned long task_ratelimit,
|
|
|
|
unsigned long dirty_ratelimit,
|
|
|
|
int *nr_dirtied_pause)
|
2011-06-12 01:21:43 +00:00
|
|
|
{
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
long hi = ilog2(bdi->avg_write_bandwidth);
|
|
|
|
long lo = ilog2(bdi->dirty_ratelimit);
|
|
|
|
long t; /* target pause */
|
|
|
|
long pause; /* estimated next pause */
|
|
|
|
int pages; /* target nr_dirtied_pause */
|
2011-06-12 01:21:43 +00:00
|
|
|
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
/* target for 10ms pause on 1-dd case */
|
|
|
|
t = max(1, HZ / 100);
|
2011-06-12 01:21:43 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Scale up pause time for concurrent dirtiers in order to reduce CPU
|
|
|
|
* overheads.
|
|
|
|
*
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
* (N * 10ms) on 2^N concurrent tasks.
|
2011-06-12 01:21:43 +00:00
|
|
|
*/
|
|
|
|
if (hi > lo)
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
t += (hi - lo) * (10 * HZ) / 1024;
|
2011-06-12 01:21:43 +00:00
|
|
|
|
|
|
|
/*
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
* This is a bit convoluted. We try to base the next nr_dirtied_pause
|
|
|
|
* on the much more stable dirty_ratelimit. However the next pause time
|
|
|
|
* will be computed based on task_ratelimit and the two rate limits may
|
|
|
|
* depart considerably at some time. Especially if task_ratelimit goes
|
|
|
|
* below dirty_ratelimit/2 and the target pause is max_pause, the next
|
|
|
|
* pause time will be max_pause*2 _trimmed down_ to max_pause. As a
|
|
|
|
* result task_ratelimit won't be executed faithfully, which could
|
|
|
|
* eventually bring down dirty_ratelimit.
|
2011-06-12 01:21:43 +00:00
|
|
|
*
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
* We apply two rules to fix it up:
|
|
|
|
* 1) try to estimate the next pause time and if necessary, use a lower
|
|
|
|
* nr_dirtied_pause so as not to exceed max_pause. When this happens,
|
|
|
|
* nr_dirtied_pause will be "dancing" with task_ratelimit.
|
|
|
|
* 2) limit the target pause time to max_pause/2, so that the normal
|
|
|
|
* small fluctuations of task_ratelimit won't trigger rule (1) and
|
|
|
|
* nr_dirtied_pause will remain as stable as dirty_ratelimit.
|
2011-06-12 01:21:43 +00:00
|
|
|
*/
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
t = min(t, 1 + max_pause / 2);
|
|
|
|
pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
|
2011-06-12 01:21:43 +00:00
|
|
|
|
|
|
|
/*
|
2011-12-06 19:17:17 +00:00
|
|
|
* Tiny nr_dirtied_pause is found to hurt I/O performance in the test
|
|
|
|
* case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
|
|
|
|
* When the 16 consecutive reads are often interrupted by some dirty
|
|
|
|
* throttling pause during the async writes, cfq will go into idles
|
|
|
|
* (deadline is fine). So push nr_dirtied_pause as high as possible
|
|
|
|
* until reaches DIRTY_POLL_THRESH=32 pages.
|
2011-06-12 01:21:43 +00:00
|
|
|
*/
|
2011-12-06 19:17:17 +00:00
|
|
|
if (pages < DIRTY_POLL_THRESH) {
|
|
|
|
t = max_pause;
|
|
|
|
pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
|
|
|
|
if (pages > DIRTY_POLL_THRESH) {
|
|
|
|
pages = DIRTY_POLL_THRESH;
|
|
|
|
t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
pause = HZ * pages / (task_ratelimit + 1);
|
|
|
|
if (pause > max_pause) {
|
|
|
|
t = max_pause;
|
|
|
|
pages = task_ratelimit * t / roundup_pow_of_two(HZ);
|
|
|
|
}
|
2011-06-12 01:21:43 +00:00
|
|
|
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
*nr_dirtied_pause = pages;
|
2011-06-12 01:21:43 +00:00
|
|
|
/*
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
* The minimal pause time will normally be half the target pause time.
|
2011-06-12 01:21:43 +00:00
|
|
|
*/
|
2011-12-06 19:17:17 +00:00
|
|
|
return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
|
2011-06-12 01:21:43 +00:00
|
|
|
}
|
|
|
|
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
static inline void bdi_dirty_limits(struct backing_dev_info *bdi,
|
|
|
|
unsigned long dirty_thresh,
|
|
|
|
unsigned long background_thresh,
|
|
|
|
unsigned long *bdi_dirty,
|
|
|
|
unsigned long *bdi_thresh,
|
|
|
|
unsigned long *bdi_bg_thresh)
|
|
|
|
{
|
|
|
|
unsigned long bdi_reclaimable;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* bdi_thresh is not treated as some limiting factor as
|
|
|
|
* dirty_thresh, due to reasons
|
|
|
|
* - in JBOD setup, bdi_thresh can fluctuate a lot
|
|
|
|
* - in a system with HDD and USB key, the USB key may somehow
|
|
|
|
* go into state (bdi_dirty >> bdi_thresh) either because
|
|
|
|
* bdi_dirty starts high, or because bdi_thresh drops low.
|
|
|
|
* In this case we don't want to hard throttle the USB key
|
|
|
|
* dirtiers for 100 seconds until bdi_dirty drops under
|
|
|
|
* bdi_thresh. Instead the auxiliary bdi control line in
|
|
|
|
* bdi_position_ratio() will let the dirtier task progress
|
|
|
|
* at some rate <= (write_bw / 2) for bringing down bdi_dirty.
|
|
|
|
*/
|
|
|
|
*bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
|
|
|
|
|
|
|
|
if (bdi_bg_thresh)
|
2014-07-30 23:08:21 +00:00
|
|
|
*bdi_bg_thresh = dirty_thresh ? div_u64((u64)*bdi_thresh *
|
|
|
|
background_thresh,
|
|
|
|
dirty_thresh) : 0;
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* In order to avoid the stacked BDI deadlock we need
|
|
|
|
* to ensure we accurately count the 'dirty' pages when
|
|
|
|
* the threshold is low.
|
|
|
|
*
|
|
|
|
* Otherwise it would be possible to get thresh+n pages
|
|
|
|
* reported dirty, even though there are thresh-m pages
|
|
|
|
* actually dirty; with m+n sitting in the percpu
|
|
|
|
* deltas.
|
|
|
|
*/
|
|
|
|
if (*bdi_thresh < 2 * bdi_stat_error(bdi)) {
|
|
|
|
bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
|
|
|
|
*bdi_dirty = bdi_reclaimable +
|
|
|
|
bdi_stat_sum(bdi, BDI_WRITEBACK);
|
|
|
|
} else {
|
|
|
|
bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
|
|
|
|
*bdi_dirty = bdi_reclaimable +
|
|
|
|
bdi_stat(bdi, BDI_WRITEBACK);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* balance_dirty_pages() must be called by processes which are generating dirty
|
|
|
|
* data. It looks at the number of dirty pages in the machine and will force
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
* the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
|
2009-09-23 17:37:09 +00:00
|
|
|
* If we're over `background_thresh' then the writeback threads are woken to
|
|
|
|
* perform some writeout.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2009-09-23 13:56:00 +00:00
|
|
|
static void balance_dirty_pages(struct address_space *mapping,
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
unsigned long pages_dirtied)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
|
2010-09-12 19:34:05 +00:00
|
|
|
unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
|
2009-01-06 22:39:29 +00:00
|
|
|
unsigned long background_thresh;
|
|
|
|
unsigned long dirty_thresh;
|
2011-06-12 01:25:42 +00:00
|
|
|
long period;
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
long pause;
|
|
|
|
long max_pause;
|
|
|
|
long min_pause;
|
|
|
|
int nr_dirtied_pause;
|
2010-08-11 21:17:37 +00:00
|
|
|
bool dirty_exceeded = false;
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
unsigned long task_ratelimit;
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
unsigned long dirty_ratelimit;
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
unsigned long pos_ratio;
|
2005-04-16 22:20:36 +00:00
|
|
|
struct backing_dev_info *bdi = mapping->backing_dev_info;
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
|
2010-08-29 17:22:30 +00:00
|
|
|
unsigned long start_time = jiffies;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
for (;;) {
|
2011-06-12 01:25:42 +00:00
|
|
|
unsigned long now = jiffies;
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
unsigned long uninitialized_var(bdi_thresh);
|
|
|
|
unsigned long thresh;
|
|
|
|
unsigned long uninitialized_var(bdi_dirty);
|
|
|
|
unsigned long dirty;
|
|
|
|
unsigned long bg_thresh;
|
2011-06-12 01:25:42 +00:00
|
|
|
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
/*
|
|
|
|
* Unstable writes are a feature of certain networked
|
|
|
|
* filesystems (i.e. NFS) in which data may have been
|
|
|
|
* written to the server's write cache, but has not yet
|
|
|
|
* been flushed to permanent storage.
|
|
|
|
*/
|
2007-11-15 00:59:15 +00:00
|
|
|
nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
|
|
|
|
global_page_state(NR_UNSTABLE_NFS);
|
2010-09-12 19:34:05 +00:00
|
|
|
nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
|
2007-11-15 00:59:15 +00:00
|
|
|
|
2010-08-11 21:17:39 +00:00
|
|
|
global_dirty_limits(&background_thresh, &dirty_thresh);
|
|
|
|
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
if (unlikely(strictlimit)) {
|
|
|
|
bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
|
|
|
|
&bdi_dirty, &bdi_thresh, &bg_thresh);
|
|
|
|
|
|
|
|
dirty = bdi_dirty;
|
|
|
|
thresh = bdi_thresh;
|
|
|
|
} else {
|
|
|
|
dirty = nr_dirty;
|
|
|
|
thresh = dirty_thresh;
|
|
|
|
bg_thresh = background_thresh;
|
|
|
|
}
|
|
|
|
|
2010-08-11 21:17:39 +00:00
|
|
|
/*
|
|
|
|
* Throttle it only when the background writeback cannot
|
|
|
|
* catch-up. This avoids (excessively) small writeouts
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
* when the bdi limits are ramping up in case of !strictlimit.
|
|
|
|
*
|
|
|
|
* In strictlimit case make decision based on the bdi counters
|
|
|
|
* and limits. Small writeouts when the bdi limits are ramping
|
|
|
|
* up are the price we consciously pay for strictlimit-ing.
|
2010-08-11 21:17:39 +00:00
|
|
|
*/
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) {
|
2011-06-12 01:25:42 +00:00
|
|
|
current->dirty_paused_when = now;
|
|
|
|
current->nr_dirtied = 0;
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
current->nr_dirtied_pause =
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
dirty_poll_interval(dirty, thresh);
|
2010-08-11 21:17:39 +00:00
|
|
|
break;
|
2011-06-12 01:25:42 +00:00
|
|
|
}
|
2010-08-11 21:17:39 +00:00
|
|
|
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
if (unlikely(!writeback_in_progress(bdi)))
|
|
|
|
bdi_start_background_writeback(bdi);
|
|
|
|
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
if (!strictlimit)
|
|
|
|
bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
|
|
|
|
&bdi_dirty, &bdi_thresh, NULL);
|
2007-11-15 00:59:15 +00:00
|
|
|
|
2011-12-04 03:26:01 +00:00
|
|
|
dirty_exceeded = (bdi_dirty > bdi_thresh) &&
|
mm/page-writeback.c: add strictlimit feature
The feature prevents mistrusted filesystems (ie: FUSE mounts created by
unprivileged users) to grow a large number of dirty pages before
throttling. For such filesystems balance_dirty_pages always check bdi
counters against bdi limits. I.e. even if global "nr_dirty" is under
"freerun", it's not allowed to skip bdi checks. The only use case for now
is fuse: it sets bdi max_ratio to 1% by default and system administrators
are supposed to expect that this limit won't be exceeded.
The feature is on if a BDI is marked by BDI_CAP_STRICTLIMIT flag. A
filesystem may set the flag when it initializes its BDI.
The problematic scenario comes from the fact that nobody pays attention to
the NR_WRITEBACK_TEMP counter (i.e. number of pages under fuse
writeback). The implementation of fuse writeback releases original page
(by calling end_page_writeback) almost immediately. A fuse request queued
for real processing bears a copy of original page. Hence, if userspace
fuse daemon doesn't finalize write requests in timely manner, an
aggressive mmap writer can pollute virtually all memory by those temporary
fuse page copies. They are carefully accounted in NR_WRITEBACK_TEMP, but
nobody cares.
To make further explanations shorter, let me use "NR_WRITEBACK_TEMP
problem" as a shortcut for "a possibility of uncontrolled grow of amount
of RAM consumed by temporary pages allocated by kernel fuse to process
writeback".
The problem was very easy to reproduce. There is a trivial example
filesystem implementation in fuse userspace distribution: fusexmp_fh.c. I
added "sleep(1);" to the write methods, then recompiled and mounted it.
Then created a huge file on the mount point and run a simple program which
mmap-ed the file to a memory region, then wrote a data to the region. An
hour later I observed almost all RAM consumed by fuse writeback. Since
then some unrelated changes in kernel fuse made it more difficult to
reproduce, but it is still possible now.
Putting this theoretical happens-in-the-lab thing aside, there is another
thing that really hurts real world (FUSE) users. This is write-through
page cache policy FUSE currently uses. I.e. handling write(2), kernel
fuse populates page cache and flushes user data to the server
synchronously. This is excessively suboptimal. Pavel Emelyanov's patches
("writeback cache policy") solve the problem, but they also make resolving
NR_WRITEBACK_TEMP problem absolutely necessary. Otherwise, simply copying
a huge file to a fuse mount would result in memory starvation. Miklos,
the maintainer of FUSE, believes strictlimit feature the way to go.
And eventually putting FUSE topics aside, there is one more use-case for
strictlimit feature. Using a slow USB stick (mass storage) in a machine
with huge amount of RAM installed is a well-known pain. Let's make simple
computations. Assuming 64GB of RAM installed, existing implementation of
balance_dirty_pages will start throttling only after 9.6GB of RAM becomes
dirty (freerun == 15% of total RAM). So, the command "cp 9GB_file
/media/my-usb-storage/" may return in a few seconds, but subsequent
"umount /media/my-usb-storage/" will take more than two hours if effective
throughput of the storage is, to say, 1MB/sec.
After inclusion of strictlimit feature, it will be trivial to add a knob
(e.g. /sys/devices/virtual/bdi/x:y/strictlimit) to enable it on demand.
Manually or via udev rule. May be I'm wrong, but it seems to be quite a
natural desire to limit the amount of dirty memory for some devices we are
not fully trust (in the sense of sustainable throughput).
[akpm@linux-foundation.org: fix warning in page-writeback.c]
Signed-off-by: Maxim Patlasov <MPatlasov@parallels.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:22:46 +00:00
|
|
|
((nr_dirty > dirty_thresh) || strictlimit);
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
if (dirty_exceeded && !bdi->dirty_exceeded)
|
2007-10-17 06:25:50 +00:00
|
|
|
bdi->dirty_exceeded = 1;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2011-10-04 02:46:17 +00:00
|
|
|
bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
|
|
|
|
nr_dirty, bdi_thresh, bdi_dirty,
|
|
|
|
start_time);
|
2010-08-29 17:22:30 +00:00
|
|
|
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
dirty_ratelimit = bdi->dirty_ratelimit;
|
|
|
|
pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
|
|
|
|
background_thresh, nr_dirty,
|
|
|
|
bdi_thresh, bdi_dirty);
|
2011-11-07 11:19:28 +00:00
|
|
|
task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
|
|
|
|
RATELIMIT_CALC_SHIFT;
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
max_pause = bdi_max_pause(bdi, bdi_dirty);
|
|
|
|
min_pause = bdi_min_pause(bdi, max_pause,
|
|
|
|
task_ratelimit, dirty_ratelimit,
|
|
|
|
&nr_dirtied_pause);
|
|
|
|
|
2011-11-07 11:19:28 +00:00
|
|
|
if (unlikely(task_ratelimit == 0)) {
|
2011-06-12 01:25:42 +00:00
|
|
|
period = max_pause;
|
2011-06-12 01:21:43 +00:00
|
|
|
pause = max_pause;
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
goto pause;
|
2007-10-17 06:25:50 +00:00
|
|
|
}
|
2011-06-12 01:25:42 +00:00
|
|
|
period = HZ * pages_dirtied / task_ratelimit;
|
|
|
|
pause = period;
|
|
|
|
if (current->dirty_paused_when)
|
|
|
|
pause -= now - current->dirty_paused_when;
|
|
|
|
/*
|
|
|
|
* For less than 1s think time (ext3/4 may block the dirtier
|
|
|
|
* for up to 800ms from time to time on 1-HDD; so does xfs,
|
|
|
|
* however at much less frequency), try to compensate it in
|
|
|
|
* future periods by updating the virtual time; otherwise just
|
|
|
|
* do a reset, as it may be a light dirtier.
|
|
|
|
*/
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
if (pause < min_pause) {
|
2010-08-30 05:33:20 +00:00
|
|
|
trace_balance_dirty_pages(bdi,
|
|
|
|
dirty_thresh,
|
|
|
|
background_thresh,
|
|
|
|
nr_dirty,
|
|
|
|
bdi_thresh,
|
|
|
|
bdi_dirty,
|
|
|
|
dirty_ratelimit,
|
|
|
|
task_ratelimit,
|
|
|
|
pages_dirtied,
|
2011-06-12 01:25:42 +00:00
|
|
|
period,
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
min(pause, 0L),
|
2010-08-30 05:33:20 +00:00
|
|
|
start_time);
|
2011-06-12 01:25:42 +00:00
|
|
|
if (pause < -HZ) {
|
|
|
|
current->dirty_paused_when = now;
|
|
|
|
current->nr_dirtied = 0;
|
|
|
|
} else if (period) {
|
|
|
|
current->dirty_paused_when += period;
|
|
|
|
current->nr_dirtied = 0;
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
} else if (current->nr_dirtied_pause <= pages_dirtied)
|
|
|
|
current->nr_dirtied_pause += pages_dirtied;
|
2011-06-12 01:32:32 +00:00
|
|
|
break;
|
2007-10-17 06:25:50 +00:00
|
|
|
}
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
if (unlikely(pause > max_pause)) {
|
|
|
|
/* for occasional dropped task_ratelimit */
|
|
|
|
now += min(pause - max_pause, max_pause);
|
|
|
|
pause = max_pause;
|
|
|
|
}
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
|
|
|
|
pause:
|
2010-08-30 05:33:20 +00:00
|
|
|
trace_balance_dirty_pages(bdi,
|
|
|
|
dirty_thresh,
|
|
|
|
background_thresh,
|
|
|
|
nr_dirty,
|
|
|
|
bdi_thresh,
|
|
|
|
bdi_dirty,
|
|
|
|
dirty_ratelimit,
|
|
|
|
task_ratelimit,
|
|
|
|
pages_dirtied,
|
2011-06-12 01:25:42 +00:00
|
|
|
period,
|
2010-08-30 05:33:20 +00:00
|
|
|
pause,
|
|
|
|
start_time);
|
2011-11-16 11:34:48 +00:00
|
|
|
__set_current_state(TASK_KILLABLE);
|
2009-10-09 10:40:42 +00:00
|
|
|
io_schedule_timeout(pause);
|
2009-09-17 17:59:14 +00:00
|
|
|
|
2011-06-12 01:25:42 +00:00
|
|
|
current->dirty_paused_when = now + pause;
|
|
|
|
current->nr_dirtied = 0;
|
writeback: max, min and target dirty pause time
Control the pause time and the call intervals to balance_dirty_pages()
with three parameters:
1) max_pause, limited by bdi_dirty and MAX_PAUSE
2) the target pause time, grows with the number of dd tasks
and is normally limited by max_pause/2
3) the minimal pause, set to half the target pause
and is used to skip short sleeps and accumulate them into bigger ones
The typical behaviors after patch:
- if ever task_ratelimit is far below dirty_ratelimit, the pause time
will remain constant at max_pause and nr_dirtied_pause will be
fluctuating with task_ratelimit
- in the normal cases, nr_dirtied_pause will remain stable (keep in the
same pace with dirty_ratelimit) and the pause time will be fluctuating
with task_ratelimit
In summary, someone has to fluctuate with task_ratelimit, because
task_ratelimit = nr_dirtied_pause / pause
We normally prefer a stable nr_dirtied_pause, until reaching max_pause.
The notable behavior changes are:
- in stable workloads, there will no longer be sudden big trajectory
switching of nr_dirtied_pause as concerned by Peter. It will be as
smooth as dirty_ratelimit and changing proportionally with it (as
always, assuming bdi bandwidth does not fluctuate across 2^N lines,
otherwise nr_dirtied_pause will show up in 2+ parallel trajectories)
- in the rare cases when something keeps task_ratelimit far below
dirty_ratelimit, the smoothness can no longer be retained and
nr_dirtied_pause will be "dancing" with task_ratelimit. This fixes a
(not that destructive but still not good) bug that
dirty_ratelimit gets brought down undesirably
<= balanced_dirty_ratelimit is under estimated
<= weakly executed task_ratelimit
<= pause goes too large and gets trimmed down to max_pause
<= nr_dirtied_pause (based on dirty_ratelimit) is set too large
<= dirty_ratelimit being much larger than task_ratelimit
- introduce min_pause to avoid small pause sleeps
- when pause is trimmed down to max_pause, try to compensate it at the
next pause time
The "refactor" type of changes are:
The max_pause equation is slightly transformed to make it slightly more
efficient.
We now scale target_pause by (N * 10ms) on 2^N concurrent tasks, which
is effectively equal to the original scaling max_pause by (N * 20ms)
because the original code does implicit target_pause ~= max_pause / 2.
Based on the same implicit ratio, target_pause starts with 10ms on 1 dd.
CC: Jan Kara <jack@suse.cz>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-11-30 17:08:55 +00:00
|
|
|
current->nr_dirtied_pause = nr_dirtied_pause;
|
2011-06-12 01:25:42 +00:00
|
|
|
|
2011-06-20 04:18:42 +00:00
|
|
|
/*
|
2011-11-14 01:47:32 +00:00
|
|
|
* This is typically equal to (nr_dirty < dirty_thresh) and can
|
|
|
|
* also keep "1000+ dd on a slow USB stick" under control.
|
2011-06-20 04:18:42 +00:00
|
|
|
*/
|
2011-11-14 01:47:32 +00:00
|
|
|
if (task_ratelimit)
|
2011-06-20 04:18:42 +00:00
|
|
|
break;
|
2011-11-16 11:34:48 +00:00
|
|
|
|
2011-12-02 16:21:33 +00:00
|
|
|
/*
|
|
|
|
* In the case of an unresponding NFS server and the NFS dirty
|
|
|
|
* pages exceeds dirty_thresh, give the other good bdi's a pipe
|
|
|
|
* to go through, so that tasks on them still remain responsive.
|
|
|
|
*
|
|
|
|
* In theory 1 page is enough to keep the comsumer-producer
|
|
|
|
* pipe going: the flusher cleans 1 page => the task dirties 1
|
|
|
|
* more page. However bdi_dirty has accounting errors. So use
|
|
|
|
* the larger and more IO friendly bdi_stat_error.
|
|
|
|
*/
|
|
|
|
if (bdi_dirty <= bdi_stat_error(bdi))
|
|
|
|
break;
|
|
|
|
|
2011-11-16 11:34:48 +00:00
|
|
|
if (fatal_signal_pending(current))
|
|
|
|
break;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
if (!dirty_exceeded && bdi->dirty_exceeded)
|
2007-10-17 06:25:50 +00:00
|
|
|
bdi->dirty_exceeded = 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
if (writeback_in_progress(bdi))
|
2009-09-23 17:37:09 +00:00
|
|
|
return;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* In laptop mode, we wait until hitting the higher threshold before
|
|
|
|
* starting background writeout, and then write out all the way down
|
|
|
|
* to the lower threshold. So slow writers cause minimal disk activity.
|
|
|
|
*
|
|
|
|
* In normal mode, we start background writeout at the lower
|
|
|
|
* background_thresh, to keep the amount of dirty memory low.
|
|
|
|
*/
|
writeback: IO-less balance_dirty_pages()
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>
2010-08-28 00:45:12 +00:00
|
|
|
if (laptop_mode)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (nr_reclaimable > background_thresh)
|
2010-06-08 16:15:15 +00:00
|
|
|
bdi_start_background_writeback(bdi);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2014-04-07 22:37:51 +00:00
|
|
|
void set_page_dirty_balance(struct page *page)
|
2006-09-26 06:30:58 +00:00
|
|
|
{
|
2014-04-07 22:37:51 +00:00
|
|
|
if (set_page_dirty(page)) {
|
2006-09-26 06:30:58 +00:00
|
|
|
struct address_space *mapping = page_mapping(page);
|
|
|
|
|
|
|
|
if (mapping)
|
|
|
|
balance_dirty_pages_ratelimited(mapping);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
static DEFINE_PER_CPU(int, bdp_ratelimits);
|
2009-06-24 06:13:48 +00:00
|
|
|
|
2011-04-05 19:21:19 +00:00
|
|
|
/*
|
|
|
|
* Normal tasks are throttled by
|
|
|
|
* loop {
|
|
|
|
* dirty tsk->nr_dirtied_pause pages;
|
|
|
|
* take a snap in balance_dirty_pages();
|
|
|
|
* }
|
|
|
|
* However there is a worst case. If every task exit immediately when dirtied
|
|
|
|
* (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
|
|
|
|
* called to throttle the page dirties. The solution is to save the not yet
|
|
|
|
* throttled page dirties in dirty_throttle_leaks on task exit and charge them
|
|
|
|
* randomly into the running tasks. This works well for the above worst case,
|
|
|
|
* as the new task will pick up and accumulate the old task's leaked dirty
|
|
|
|
* count and eventually get throttled.
|
|
|
|
*/
|
|
|
|
DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/**
|
2012-12-12 00:00:21 +00:00
|
|
|
* balance_dirty_pages_ratelimited - balance dirty memory state
|
2005-05-01 15:59:26 +00:00
|
|
|
* @mapping: address_space which was dirtied
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* Processes which are dirtying memory should call in here once for each page
|
|
|
|
* which was newly dirtied. The function will periodically check the system's
|
|
|
|
* dirty state and will initiate writeback if needed.
|
|
|
|
*
|
|
|
|
* On really big machines, get_writeback_state is expensive, so try to avoid
|
|
|
|
* calling it too often (ratelimiting). But once we're over the dirty memory
|
|
|
|
* limit we decrease the ratelimiting by a lot, to prevent individual processes
|
|
|
|
* from overshooting the limit by (ratelimit_pages) each.
|
|
|
|
*/
|
2012-12-12 00:00:21 +00:00
|
|
|
void balance_dirty_pages_ratelimited(struct address_space *mapping)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2011-06-11 23:53:57 +00:00
|
|
|
struct backing_dev_info *bdi = mapping->backing_dev_info;
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
int ratelimit;
|
|
|
|
int *p;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2011-06-11 23:53:57 +00:00
|
|
|
if (!bdi_cap_account_dirty(bdi))
|
|
|
|
return;
|
|
|
|
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
ratelimit = current->nr_dirtied_pause;
|
|
|
|
if (bdi->dirty_exceeded)
|
|
|
|
ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
|
|
|
|
|
|
|
|
preempt_disable();
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
* This prevents one CPU to accumulate too many dirtied pages without
|
|
|
|
* calling into balance_dirty_pages(), which can happen when there are
|
|
|
|
* 1000+ tasks, all of them start dirtying pages at exactly the same
|
|
|
|
* time, hence all honoured too large initial task->nr_dirtied_pause.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2014-06-04 23:07:56 +00:00
|
|
|
p = this_cpu_ptr(&bdp_ratelimits);
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
if (unlikely(current->nr_dirtied >= ratelimit))
|
2006-03-24 11:18:10 +00:00
|
|
|
*p = 0;
|
2011-04-14 13:52:37 +00:00
|
|
|
else if (unlikely(*p >= ratelimit_pages)) {
|
|
|
|
*p = 0;
|
|
|
|
ratelimit = 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2011-04-05 19:21:19 +00:00
|
|
|
/*
|
|
|
|
* Pick up the dirtied pages by the exited tasks. This avoids lots of
|
|
|
|
* short-lived tasks (eg. gcc invocations in a kernel build) escaping
|
|
|
|
* the dirty throttling and livelock other long-run dirtiers.
|
|
|
|
*/
|
2014-06-04 23:07:56 +00:00
|
|
|
p = this_cpu_ptr(&dirty_throttle_leaks);
|
2011-04-05 19:21:19 +00:00
|
|
|
if (*p > 0 && current->nr_dirtied < ratelimit) {
|
2012-12-12 00:00:21 +00:00
|
|
|
unsigned long nr_pages_dirtied;
|
2011-04-05 19:21:19 +00:00
|
|
|
nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
|
|
|
|
*p -= nr_pages_dirtied;
|
|
|
|
current->nr_dirtied += nr_pages_dirtied;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2006-03-24 11:18:10 +00:00
|
|
|
preempt_enable();
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
|
|
|
|
if (unlikely(current->nr_dirtied >= ratelimit))
|
|
|
|
balance_dirty_pages(mapping, current->nr_dirtied);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2012-12-12 00:00:21 +00:00
|
|
|
EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2007-03-01 04:13:21 +00:00
|
|
|
void throttle_vm_writeout(gfp_t gfp_mask)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2009-01-06 22:39:29 +00:00
|
|
|
unsigned long background_thresh;
|
|
|
|
unsigned long dirty_thresh;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
for ( ; ; ) {
|
2010-08-11 21:17:39 +00:00
|
|
|
global_dirty_limits(&background_thresh, &dirty_thresh);
|
2012-03-21 23:34:09 +00:00
|
|
|
dirty_thresh = hard_dirty_limit(dirty_thresh);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Boost the allowable dirty threshold a bit for page
|
|
|
|
* allocators so they don't get DoS'ed by heavy writers
|
|
|
|
*/
|
|
|
|
dirty_thresh += dirty_thresh / 10; /* wheeee... */
|
|
|
|
|
2006-06-30 08:55:42 +00:00
|
|
|
if (global_page_state(NR_UNSTABLE_NFS) +
|
|
|
|
global_page_state(NR_WRITEBACK) <= dirty_thresh)
|
|
|
|
break;
|
2009-07-09 12:52:32 +00:00
|
|
|
congestion_wait(BLK_RW_ASYNC, HZ/10);
|
2007-10-17 06:30:45 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* The caller might hold locks which can prevent IO completion
|
|
|
|
* or progress in the filesystem. So we cannot just sit here
|
|
|
|
* waiting for IO to complete.
|
|
|
|
*/
|
|
|
|
if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
|
|
|
|
break;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
|
|
|
|
*/
|
2014-06-06 21:38:09 +00:00
|
|
|
int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
|
2009-09-23 22:57:19 +00:00
|
|
|
void __user *buffer, size_t *length, loff_t *ppos)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2009-09-23 22:57:19 +00:00
|
|
|
proc_dointvec(table, write, buffer, length, ppos);
|
2005-04-16 22:20:36 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2010-05-20 07:18:47 +00:00
|
|
|
#ifdef CONFIG_BLOCK
|
2010-04-06 12:25:14 +00:00
|
|
|
void laptop_mode_timer_fn(unsigned long data)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2010-04-06 12:25:14 +00:00
|
|
|
struct request_queue *q = (struct request_queue *)data;
|
|
|
|
int nr_pages = global_page_state(NR_FILE_DIRTY) +
|
|
|
|
global_page_state(NR_UNSTABLE_NFS);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2010-04-06 12:25:14 +00:00
|
|
|
/*
|
|
|
|
* We want to write everything out, not just down to the dirty
|
|
|
|
* threshold
|
|
|
|
*/
|
|
|
|
if (bdi_has_dirty_io(&q->backing_dev_info))
|
2011-10-08 03:54:10 +00:00
|
|
|
bdi_start_writeback(&q->backing_dev_info, nr_pages,
|
|
|
|
WB_REASON_LAPTOP_TIMER);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We've spun up the disk and we're in laptop mode: schedule writeback
|
|
|
|
* of all dirty data a few seconds from now. If the flush is already scheduled
|
|
|
|
* then push it back - the user is still using the disk.
|
|
|
|
*/
|
2010-04-06 12:25:14 +00:00
|
|
|
void laptop_io_completion(struct backing_dev_info *info)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2010-04-06 12:25:14 +00:00
|
|
|
mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We're in laptop mode and we've just synced. The sync's writes will have
|
|
|
|
* caused another writeback to be scheduled by laptop_io_completion.
|
|
|
|
* Nothing needs to be written back anymore, so we unschedule the writeback.
|
|
|
|
*/
|
|
|
|
void laptop_sync_completion(void)
|
|
|
|
{
|
2010-04-06 12:25:14 +00:00
|
|
|
struct backing_dev_info *bdi;
|
|
|
|
|
|
|
|
rcu_read_lock();
|
|
|
|
|
|
|
|
list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
|
|
|
|
del_timer(&bdi->laptop_mode_wb_timer);
|
|
|
|
|
|
|
|
rcu_read_unlock();
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2010-05-20 07:18:47 +00:00
|
|
|
#endif
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* If ratelimit_pages is too high then we can get into dirty-data overload
|
|
|
|
* if a large number of processes all perform writes at the same time.
|
|
|
|
* If it is too low then SMP machines will call the (expensive)
|
|
|
|
* get_writeback_state too often.
|
|
|
|
*
|
|
|
|
* Here we set ratelimit_pages to a level which ensures that when all CPUs are
|
|
|
|
* dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
* thresholds.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
|
|
|
|
2006-09-29 09:01:25 +00:00
|
|
|
void writeback_set_ratelimit(void)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
unsigned long background_thresh;
|
|
|
|
unsigned long dirty_thresh;
|
|
|
|
global_dirty_limits(&background_thresh, &dirty_thresh);
|
2012-05-06 05:21:42 +00:00
|
|
|
global_dirty_limit = dirty_thresh;
|
writeback: per task dirty rate limit
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>
2011-06-12 00:10:12 +00:00
|
|
|
ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
|
2005-04-16 22:20:36 +00:00
|
|
|
if (ratelimit_pages < 16)
|
|
|
|
ratelimit_pages = 16;
|
|
|
|
}
|
|
|
|
|
2013-06-19 18:53:51 +00:00
|
|
|
static int
|
2012-09-28 12:27:49 +00:00
|
|
|
ratelimit_handler(struct notifier_block *self, unsigned long action,
|
|
|
|
void *hcpu)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2012-09-28 12:27:49 +00:00
|
|
|
|
|
|
|
switch (action & ~CPU_TASKS_FROZEN) {
|
|
|
|
case CPU_ONLINE:
|
|
|
|
case CPU_DEAD:
|
|
|
|
writeback_set_ratelimit();
|
|
|
|
return NOTIFY_OK;
|
|
|
|
default:
|
|
|
|
return NOTIFY_DONE;
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2013-06-19 18:53:51 +00:00
|
|
|
static struct notifier_block ratelimit_nb = {
|
2005-04-16 22:20:36 +00:00
|
|
|
.notifier_call = ratelimit_handler,
|
|
|
|
.next = NULL,
|
|
|
|
};
|
|
|
|
|
|
|
|
/*
|
2007-01-30 00:37:38 +00:00
|
|
|
* Called early on to tune the page writeback dirty limits.
|
|
|
|
*
|
|
|
|
* We used to scale dirty pages according to how total memory
|
|
|
|
* related to pages that could be allocated for buffers (by
|
|
|
|
* comparing nr_free_buffer_pages() to vm_total_pages.
|
|
|
|
*
|
|
|
|
* However, that was when we used "dirty_ratio" to scale with
|
|
|
|
* all memory, and we don't do that any more. "dirty_ratio"
|
|
|
|
* is now applied to total non-HIGHPAGE memory (by subtracting
|
|
|
|
* totalhigh_pages from vm_total_pages), and as such we can't
|
|
|
|
* get into the old insane situation any more where we had
|
|
|
|
* large amounts of dirty pages compared to a small amount of
|
|
|
|
* non-HIGHMEM memory.
|
|
|
|
*
|
|
|
|
* But we might still want to scale the dirty_ratio by how
|
|
|
|
* much memory the box has..
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
|
|
|
void __init page_writeback_init(void)
|
|
|
|
{
|
2006-09-29 09:01:25 +00:00
|
|
|
writeback_set_ratelimit();
|
2005-04-16 22:20:36 +00:00
|
|
|
register_cpu_notifier(&ratelimit_nb);
|
2007-10-17 06:25:50 +00:00
|
|
|
|
2012-05-24 16:59:11 +00:00
|
|
|
fprop_global_init(&writeout_completions);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2010-08-10 00:19:12 +00:00
|
|
|
/**
|
|
|
|
* tag_pages_for_writeback - tag pages to be written by write_cache_pages
|
|
|
|
* @mapping: address space structure to write
|
|
|
|
* @start: starting page index
|
|
|
|
* @end: ending page index (inclusive)
|
|
|
|
*
|
|
|
|
* This function scans the page range from @start to @end (inclusive) and tags
|
|
|
|
* all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
|
|
|
|
* that write_cache_pages (or whoever calls this function) will then use
|
|
|
|
* TOWRITE tag to identify pages eligible for writeback. This mechanism is
|
|
|
|
* used to avoid livelocking of writeback by a process steadily creating new
|
|
|
|
* dirty pages in the file (thus it is important for this function to be quick
|
|
|
|
* so that it can tag pages faster than a dirtying process can create them).
|
|
|
|
*/
|
|
|
|
/*
|
|
|
|
* We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
|
|
|
|
*/
|
|
|
|
void tag_pages_for_writeback(struct address_space *mapping,
|
|
|
|
pgoff_t start, pgoff_t end)
|
|
|
|
{
|
2010-08-11 21:17:30 +00:00
|
|
|
#define WRITEBACK_TAG_BATCH 4096
|
2010-08-10 00:19:12 +00:00
|
|
|
unsigned long tagged;
|
|
|
|
|
|
|
|
do {
|
|
|
|
spin_lock_irq(&mapping->tree_lock);
|
|
|
|
tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
|
|
|
|
&start, end, WRITEBACK_TAG_BATCH,
|
|
|
|
PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
|
|
|
|
spin_unlock_irq(&mapping->tree_lock);
|
|
|
|
WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
|
|
|
|
cond_resched();
|
2010-08-19 21:13:33 +00:00
|
|
|
/* We check 'start' to handle wrapping when end == ~0UL */
|
|
|
|
} while (tagged >= WRITEBACK_TAG_BATCH && start);
|
2010-08-10 00:19:12 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(tag_pages_for_writeback);
|
|
|
|
|
2006-08-29 18:06:09 +00:00
|
|
|
/**
|
2007-05-11 05:22:51 +00:00
|
|
|
* write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
|
2006-08-29 18:06:09 +00:00
|
|
|
* @mapping: address space structure to write
|
|
|
|
* @wbc: subtract the number of written pages from *@wbc->nr_to_write
|
2007-05-11 05:22:51 +00:00
|
|
|
* @writepage: function called for each page
|
|
|
|
* @data: data passed to writepage function
|
2006-08-29 18:06:09 +00:00
|
|
|
*
|
2007-05-11 05:22:51 +00:00
|
|
|
* If a page is already under I/O, write_cache_pages() skips it, even
|
2006-08-29 18:06:09 +00:00
|
|
|
* if it's dirty. This is desirable behaviour for memory-cleaning writeback,
|
|
|
|
* but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
|
|
|
|
* and msync() need to guarantee that all the data which was dirty at the time
|
|
|
|
* the call was made get new I/O started against them. If wbc->sync_mode is
|
|
|
|
* WB_SYNC_ALL then we were called for data integrity and we must wait for
|
|
|
|
* existing IO to complete.
|
2010-08-10 00:19:12 +00:00
|
|
|
*
|
|
|
|
* To avoid livelocks (when other process dirties new pages), we first tag
|
|
|
|
* pages which should be written back with TOWRITE tag and only then start
|
|
|
|
* writing them. For data-integrity sync we have to be careful so that we do
|
|
|
|
* not miss some pages (e.g., because some other process has cleared TOWRITE
|
|
|
|
* tag we set). The rule we follow is that TOWRITE tag can be cleared only
|
|
|
|
* by the process clearing the DIRTY tag (and submitting the page for IO).
|
2006-08-29 18:06:09 +00:00
|
|
|
*/
|
2007-05-11 05:22:51 +00:00
|
|
|
int write_cache_pages(struct address_space *mapping,
|
|
|
|
struct writeback_control *wbc, writepage_t writepage,
|
|
|
|
void *data)
|
2006-08-29 18:06:09 +00:00
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
int done = 0;
|
|
|
|
struct pagevec pvec;
|
|
|
|
int nr_pages;
|
2009-01-06 22:39:04 +00:00
|
|
|
pgoff_t uninitialized_var(writeback_index);
|
2006-08-29 18:06:09 +00:00
|
|
|
pgoff_t index;
|
|
|
|
pgoff_t end; /* Inclusive */
|
2009-01-06 22:39:06 +00:00
|
|
|
pgoff_t done_index;
|
2009-01-06 22:39:04 +00:00
|
|
|
int cycled;
|
2006-08-29 18:06:09 +00:00
|
|
|
int range_whole = 0;
|
2010-08-10 00:19:12 +00:00
|
|
|
int tag;
|
2006-08-29 18:06:09 +00:00
|
|
|
|
|
|
|
pagevec_init(&pvec, 0);
|
|
|
|
if (wbc->range_cyclic) {
|
2009-01-06 22:39:04 +00:00
|
|
|
writeback_index = mapping->writeback_index; /* prev offset */
|
|
|
|
index = writeback_index;
|
|
|
|
if (index == 0)
|
|
|
|
cycled = 1;
|
|
|
|
else
|
|
|
|
cycled = 0;
|
2006-08-29 18:06:09 +00:00
|
|
|
end = -1;
|
|
|
|
} else {
|
|
|
|
index = wbc->range_start >> PAGE_CACHE_SHIFT;
|
|
|
|
end = wbc->range_end >> PAGE_CACHE_SHIFT;
|
|
|
|
if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
|
|
|
|
range_whole = 1;
|
2009-01-06 22:39:04 +00:00
|
|
|
cycled = 1; /* ignore range_cyclic tests */
|
2006-08-29 18:06:09 +00:00
|
|
|
}
|
2010-06-06 16:38:15 +00:00
|
|
|
if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
|
2010-08-10 00:19:12 +00:00
|
|
|
tag = PAGECACHE_TAG_TOWRITE;
|
|
|
|
else
|
|
|
|
tag = PAGECACHE_TAG_DIRTY;
|
2006-08-29 18:06:09 +00:00
|
|
|
retry:
|
2010-06-06 16:38:15 +00:00
|
|
|
if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
|
2010-08-10 00:19:12 +00:00
|
|
|
tag_pages_for_writeback(mapping, index, end);
|
2009-01-06 22:39:06 +00:00
|
|
|
done_index = index;
|
2009-01-06 22:39:09 +00:00
|
|
|
while (!done && (index <= end)) {
|
|
|
|
int i;
|
|
|
|
|
2010-08-10 00:19:12 +00:00
|
|
|
nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
|
2009-01-06 22:39:09 +00:00
|
|
|
min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
|
|
|
|
if (nr_pages == 0)
|
|
|
|
break;
|
2006-08-29 18:06:09 +00:00
|
|
|
|
|
|
|
for (i = 0; i < nr_pages; i++) {
|
|
|
|
struct page *page = pvec.pages[i];
|
|
|
|
|
|
|
|
/*
|
2009-01-06 22:39:11 +00:00
|
|
|
* At this point, the page may be truncated or
|
|
|
|
* invalidated (changing page->mapping to NULL), or
|
|
|
|
* even swizzled back from swapper_space to tmpfs file
|
|
|
|
* mapping. However, page->index will not change
|
|
|
|
* because we have a reference on the page.
|
2006-08-29 18:06:09 +00:00
|
|
|
*/
|
2009-01-06 22:39:11 +00:00
|
|
|
if (page->index > end) {
|
|
|
|
/*
|
|
|
|
* can't be range_cyclic (1st pass) because
|
|
|
|
* end == -1 in that case.
|
|
|
|
*/
|
|
|
|
done = 1;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
writeback: make mapping->writeback_index to point to the last written page
For range-cyclic writeback (e.g. kupdate), the writeback code sets a
continuation point of the next writeback to mapping->writeback_index which
is set the page after the last written page. This happens so that we
evenly write the whole file even if pages in it get continuously
redirtied.
However, in some cases, sequential writer is writing in the middle of the
page and it just redirties the last written page by continuing from that.
For example with an application which uses a file as a big ring buffer we
see:
[1st writeback session]
...
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898514 + 8
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898522 + 8
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898530 + 8
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898538 + 8
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898546 + 8
kworker/0:1-11 4571: block_rq_issue: 8,0 W 0 () 94898514 + 40
>> flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898554 + 8
>> flush-8:0-2743 4571: block_rq_issue: 8,0 W 0 () 94898554 + 8
[2nd writeback session after 35sec]
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94898562 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94898570 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94898578 + 8
...
kworker/0:1-11 4606: block_rq_issue: 8,0 W 0 () 94898562 + 640
kworker/0:1-11 4606: block_rq_issue: 8,0 W 0 () 94899202 + 72
...
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899962 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899970 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899978 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899986 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899994 + 8
kworker/0:1-11 4606: block_rq_issue: 8,0 W 0 () 94899962 + 40
>> flush-8:0-2743 4606: block_bio_queue: 8,0 W 94898554 + 8
>> flush-8:0-2743 4606: block_rq_issue: 8,0 W 0 () 94898554 + 8
So we seeked back to 94898554 after we wrote all the pages at the end of
the file.
This extra seek seems unnecessary. If we continue writeback from the last
written page, we can avoid it and do not cause harm to other cases. The
original intent of even writeout over the whole file is preserved and if
the page does not get redirtied pagevec_lookup_tag() just skips it.
As an exceptional case, when I/O error happens, set done_index to the next
page as the comment in the code suggests.
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-22 23:33:40 +00:00
|
|
|
done_index = page->index;
|
2009-01-06 22:39:11 +00:00
|
|
|
|
2006-08-29 18:06:09 +00:00
|
|
|
lock_page(page);
|
|
|
|
|
2009-01-06 22:39:09 +00:00
|
|
|
/*
|
|
|
|
* Page truncated or invalidated. We can freely skip it
|
|
|
|
* then, even for data integrity operations: the page
|
|
|
|
* has disappeared concurrently, so there could be no
|
|
|
|
* real expectation of this data interity operation
|
|
|
|
* even if there is now a new, dirty page at the same
|
|
|
|
* pagecache address.
|
|
|
|
*/
|
2006-08-29 18:06:09 +00:00
|
|
|
if (unlikely(page->mapping != mapping)) {
|
2009-01-06 22:39:09 +00:00
|
|
|
continue_unlock:
|
2006-08-29 18:06:09 +00:00
|
|
|
unlock_page(page);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2009-01-06 22:39:10 +00:00
|
|
|
if (!PageDirty(page)) {
|
|
|
|
/* someone wrote it for us */
|
|
|
|
goto continue_unlock;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (PageWriteback(page)) {
|
|
|
|
if (wbc->sync_mode != WB_SYNC_NONE)
|
|
|
|
wait_on_page_writeback(page);
|
|
|
|
else
|
|
|
|
goto continue_unlock;
|
|
|
|
}
|
2006-08-29 18:06:09 +00:00
|
|
|
|
2009-01-06 22:39:10 +00:00
|
|
|
BUG_ON(PageWriteback(page));
|
|
|
|
if (!clear_page_dirty_for_io(page))
|
2009-01-06 22:39:09 +00:00
|
|
|
goto continue_unlock;
|
2006-08-29 18:06:09 +00:00
|
|
|
|
2010-07-07 03:24:08 +00:00
|
|
|
trace_wbc_writepage(wbc, mapping->backing_dev_info);
|
2007-05-11 05:22:51 +00:00
|
|
|
ret = (*writepage)(page, wbc, data);
|
2009-01-06 22:39:06 +00:00
|
|
|
if (unlikely(ret)) {
|
|
|
|
if (ret == AOP_WRITEPAGE_ACTIVATE) {
|
|
|
|
unlock_page(page);
|
|
|
|
ret = 0;
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* done_index is set past this page,
|
|
|
|
* so media errors will not choke
|
|
|
|
* background writeout for the entire
|
|
|
|
* file. This has consequences for
|
|
|
|
* range_cyclic semantics (ie. it may
|
|
|
|
* not be suitable for data integrity
|
|
|
|
* writeout).
|
|
|
|
*/
|
writeback: make mapping->writeback_index to point to the last written page
For range-cyclic writeback (e.g. kupdate), the writeback code sets a
continuation point of the next writeback to mapping->writeback_index which
is set the page after the last written page. This happens so that we
evenly write the whole file even if pages in it get continuously
redirtied.
However, in some cases, sequential writer is writing in the middle of the
page and it just redirties the last written page by continuing from that.
For example with an application which uses a file as a big ring buffer we
see:
[1st writeback session]
...
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898514 + 8
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898522 + 8
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898530 + 8
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898538 + 8
flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898546 + 8
kworker/0:1-11 4571: block_rq_issue: 8,0 W 0 () 94898514 + 40
>> flush-8:0-2743 4571: block_bio_queue: 8,0 W 94898554 + 8
>> flush-8:0-2743 4571: block_rq_issue: 8,0 W 0 () 94898554 + 8
[2nd writeback session after 35sec]
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94898562 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94898570 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94898578 + 8
...
kworker/0:1-11 4606: block_rq_issue: 8,0 W 0 () 94898562 + 640
kworker/0:1-11 4606: block_rq_issue: 8,0 W 0 () 94899202 + 72
...
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899962 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899970 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899978 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899986 + 8
flush-8:0-2743 4606: block_bio_queue: 8,0 W 94899994 + 8
kworker/0:1-11 4606: block_rq_issue: 8,0 W 0 () 94899962 + 40
>> flush-8:0-2743 4606: block_bio_queue: 8,0 W 94898554 + 8
>> flush-8:0-2743 4606: block_rq_issue: 8,0 W 0 () 94898554 + 8
So we seeked back to 94898554 after we wrote all the pages at the end of
the file.
This extra seek seems unnecessary. If we continue writeback from the last
written page, we can avoid it and do not cause harm to other cases. The
original intent of even writeout over the whole file is preserved and if
the page does not get redirtied pagevec_lookup_tag() just skips it.
As an exceptional case, when I/O error happens, set done_index to the next
page as the comment in the code suggests.
Tested-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-22 23:33:40 +00:00
|
|
|
done_index = page->index + 1;
|
2009-01-06 22:39:06 +00:00
|
|
|
done = 1;
|
|
|
|
break;
|
|
|
|
}
|
2010-06-09 00:37:18 +00:00
|
|
|
}
|
2009-01-06 22:39:06 +00:00
|
|
|
|
2010-08-24 01:44:34 +00:00
|
|
|
/*
|
|
|
|
* We stop writing back only if we are not doing
|
|
|
|
* integrity sync. In case of integrity sync we have to
|
|
|
|
* keep going until we have written all the pages
|
|
|
|
* we tagged for writeback prior to entering this loop.
|
|
|
|
*/
|
|
|
|
if (--wbc->nr_to_write <= 0 &&
|
|
|
|
wbc->sync_mode == WB_SYNC_NONE) {
|
|
|
|
done = 1;
|
|
|
|
break;
|
mm: write_cache_pages integrity fix
In write_cache_pages, nr_to_write is heeded even for data-integrity syncs,
so the function will return success after writing out nr_to_write pages,
even if that was not sufficient to guarantee data integrity.
The callers tend to set it to values that could break data interity
semantics easily in practice. For example, nr_to_write can be set to
mapping->nr_pages * 2, however if a file has a single, dirty page, then
fsync is called, subsequent pages might be concurrently added and dirtied,
then write_cache_pages might writeout two of these newly dirty pages,
while not writing out the old page that should have been written out.
Fix this by ignoring nr_to_write if it is a data integrity sync.
This is a data integrity bug.
The reason this has been done in the past is to avoid stalling sync
operations behind page dirtiers.
"If a file has one dirty page at offset 1000000000000000 then someone
does an fsync() and someone else gets in first and starts madly writing
pages at offset 0, we want to write that page at 1000000000000000.
Somehow."
What we do today is return success after an arbitrary amount of pages are
written, whether or not we have provided the data-integrity semantics that
the caller has asked for. Even this doesn't actually fix all stall cases
completely: in the above situation, if the file has a huge number of pages
in pagecache (but not dirty), then mapping->nrpages is going to be huge,
even if pages are being dirtied.
This change does indeed make the possibility of long stalls lager, and
that's not a good thing, but lying about data integrity is even worse. We
have to either perform the sync, or return -ELINUXISLAME so at least the
caller knows what has happened.
There are subsequent competing approaches in the works to solve the stall
problems properly, without compromising data integrity.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:08 +00:00
|
|
|
}
|
2006-08-29 18:06:09 +00:00
|
|
|
}
|
|
|
|
pagevec_release(&pvec);
|
|
|
|
cond_resched();
|
|
|
|
}
|
2009-02-12 03:34:23 +00:00
|
|
|
if (!cycled && !done) {
|
2006-08-29 18:06:09 +00:00
|
|
|
/*
|
2009-01-06 22:39:04 +00:00
|
|
|
* range_cyclic:
|
2006-08-29 18:06:09 +00:00
|
|
|
* We hit the last page and there is more work to be done: wrap
|
|
|
|
* back to the start of the file
|
|
|
|
*/
|
2009-01-06 22:39:04 +00:00
|
|
|
cycled = 1;
|
2006-08-29 18:06:09 +00:00
|
|
|
index = 0;
|
2009-01-06 22:39:04 +00:00
|
|
|
end = writeback_index - 1;
|
2006-08-29 18:06:09 +00:00
|
|
|
goto retry;
|
|
|
|
}
|
2010-06-09 00:37:18 +00:00
|
|
|
if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
|
|
|
|
mapping->writeback_index = done_index;
|
2008-07-11 23:27:31 +00:00
|
|
|
|
2006-08-29 18:06:09 +00:00
|
|
|
return ret;
|
|
|
|
}
|
2007-05-11 05:22:51 +00:00
|
|
|
EXPORT_SYMBOL(write_cache_pages);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Function used by generic_writepages to call the real writepage
|
|
|
|
* function and set the mapping flags on error
|
|
|
|
*/
|
|
|
|
static int __writepage(struct page *page, struct writeback_control *wbc,
|
|
|
|
void *data)
|
|
|
|
{
|
|
|
|
struct address_space *mapping = data;
|
|
|
|
int ret = mapping->a_ops->writepage(page, wbc);
|
|
|
|
mapping_set_error(mapping, ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
|
|
|
|
* @mapping: address space structure to write
|
|
|
|
* @wbc: subtract the number of written pages from *@wbc->nr_to_write
|
|
|
|
*
|
|
|
|
* This is a library function, which implements the writepages()
|
|
|
|
* address_space_operation.
|
|
|
|
*/
|
|
|
|
int generic_writepages(struct address_space *mapping,
|
|
|
|
struct writeback_control *wbc)
|
|
|
|
{
|
2011-03-17 09:47:06 +00:00
|
|
|
struct blk_plug plug;
|
|
|
|
int ret;
|
|
|
|
|
2007-05-11 05:22:51 +00:00
|
|
|
/* deal with chardevs and other special file */
|
|
|
|
if (!mapping->a_ops->writepage)
|
|
|
|
return 0;
|
|
|
|
|
2011-03-17 09:47:06 +00:00
|
|
|
blk_start_plug(&plug);
|
|
|
|
ret = write_cache_pages(mapping, wbc, __writepage, mapping);
|
|
|
|
blk_finish_plug(&plug);
|
|
|
|
return ret;
|
2007-05-11 05:22:51 +00:00
|
|
|
}
|
2006-08-29 18:06:09 +00:00
|
|
|
|
|
|
|
EXPORT_SYMBOL(generic_writepages);
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
|
|
|
|
{
|
2005-11-16 23:07:01 +00:00
|
|
|
int ret;
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
if (wbc->nr_to_write <= 0)
|
|
|
|
return 0;
|
|
|
|
if (mapping->a_ops->writepages)
|
2006-09-26 06:30:57 +00:00
|
|
|
ret = mapping->a_ops->writepages(mapping, wbc);
|
2005-11-16 23:07:01 +00:00
|
|
|
else
|
|
|
|
ret = generic_writepages(mapping, wbc);
|
|
|
|
return ret;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* write_one_page - write out a single page and optionally wait on I/O
|
2005-05-01 15:59:26 +00:00
|
|
|
* @page: the page to write
|
|
|
|
* @wait: if true, wait on writeout
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* The page must be locked by the caller and will be unlocked upon return.
|
|
|
|
*
|
|
|
|
* write_one_page() returns a negative error code if I/O failed.
|
|
|
|
*/
|
|
|
|
int write_one_page(struct page *page, int wait)
|
|
|
|
{
|
|
|
|
struct address_space *mapping = page->mapping;
|
|
|
|
int ret = 0;
|
|
|
|
struct writeback_control wbc = {
|
|
|
|
.sync_mode = WB_SYNC_ALL,
|
|
|
|
.nr_to_write = 1,
|
|
|
|
};
|
|
|
|
|
|
|
|
BUG_ON(!PageLocked(page));
|
|
|
|
|
|
|
|
if (wait)
|
|
|
|
wait_on_page_writeback(page);
|
|
|
|
|
|
|
|
if (clear_page_dirty_for_io(page)) {
|
|
|
|
page_cache_get(page);
|
|
|
|
ret = mapping->a_ops->writepage(page, &wbc);
|
|
|
|
if (ret == 0 && wait) {
|
|
|
|
wait_on_page_writeback(page);
|
|
|
|
if (PageError(page))
|
|
|
|
ret = -EIO;
|
|
|
|
}
|
|
|
|
page_cache_release(page);
|
|
|
|
} else {
|
|
|
|
unlock_page(page);
|
|
|
|
}
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(write_one_page);
|
|
|
|
|
2007-02-10 09:43:15 +00:00
|
|
|
/*
|
|
|
|
* For address_spaces which do not use buffers nor write back.
|
|
|
|
*/
|
|
|
|
int __set_page_dirty_no_writeback(struct page *page)
|
|
|
|
{
|
|
|
|
if (!PageDirty(page))
|
2011-01-13 23:45:49 +00:00
|
|
|
return !TestSetPageDirty(page);
|
2007-02-10 09:43:15 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2009-03-31 22:19:39 +00:00
|
|
|
/*
|
|
|
|
* Helper function for set_page_dirty family.
|
|
|
|
* NOTE: This relies on being atomic wrt interrupts.
|
|
|
|
*/
|
|
|
|
void account_page_dirtied(struct page *page, struct address_space *mapping)
|
|
|
|
{
|
2013-01-11 21:06:37 +00:00
|
|
|
trace_writeback_dirty_page(page, mapping);
|
|
|
|
|
2009-03-31 22:19:39 +00:00
|
|
|
if (mapping_cap_account_dirty(mapping)) {
|
|
|
|
__inc_zone_page_state(page, NR_FILE_DIRTY);
|
2010-10-26 21:21:35 +00:00
|
|
|
__inc_zone_page_state(page, NR_DIRTIED);
|
2009-03-31 22:19:39 +00:00
|
|
|
__inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
|
2011-01-23 16:07:47 +00:00
|
|
|
__inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
|
2009-03-31 22:19:39 +00:00
|
|
|
task_io_account_write(PAGE_CACHE_SIZE);
|
2011-04-14 13:52:37 +00:00
|
|
|
current->nr_dirtied++;
|
|
|
|
this_cpu_inc(bdp_ratelimits);
|
2009-03-31 22:19:39 +00:00
|
|
|
}
|
|
|
|
}
|
2010-08-20 09:31:26 +00:00
|
|
|
EXPORT_SYMBOL(account_page_dirtied);
|
2009-03-31 22:19:39 +00:00
|
|
|
|
2010-10-26 21:21:33 +00:00
|
|
|
/*
|
|
|
|
* Helper function for set_page_writeback family.
|
2013-09-12 22:13:53 +00:00
|
|
|
*
|
|
|
|
* The caller must hold mem_cgroup_begin/end_update_page_stat() lock
|
|
|
|
* while calling this function.
|
|
|
|
* See test_set_page_writeback for example.
|
|
|
|
*
|
2010-10-26 21:21:33 +00:00
|
|
|
* NOTE: Unlike account_page_dirtied this does not rely on being atomic
|
|
|
|
* wrt interrupts.
|
|
|
|
*/
|
|
|
|
void account_page_writeback(struct page *page)
|
|
|
|
{
|
2013-09-12 22:13:53 +00:00
|
|
|
mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
|
2010-10-26 21:21:33 +00:00
|
|
|
inc_zone_page_state(page, NR_WRITEBACK);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(account_page_writeback);
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* For address_spaces which do not use buffers. Just tag the page as dirty in
|
|
|
|
* its radix tree.
|
|
|
|
*
|
|
|
|
* This is also used when a single buffer is being dirtied: we want to set the
|
|
|
|
* page dirty in that case, but not all the buffers. This is a "bottom-up"
|
|
|
|
* dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
|
|
|
|
*
|
|
|
|
* Most callers have locked the page, which pins the address_space in memory.
|
|
|
|
* But zap_pte_range() does not lock the page, however in that case the
|
|
|
|
* mapping is pinned by the vma's ->vm_file reference.
|
|
|
|
*
|
|
|
|
* We take care to handle the case where the page was truncated from the
|
2007-10-19 23:27:18 +00:00
|
|
|
* mapping by re-checking page_mapping() inside tree_lock.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
|
|
|
int __set_page_dirty_nobuffers(struct page *page)
|
|
|
|
{
|
|
|
|
if (!TestSetPageDirty(page)) {
|
|
|
|
struct address_space *mapping = page_mapping(page);
|
|
|
|
struct address_space *mapping2;
|
2014-02-06 20:04:24 +00:00
|
|
|
unsigned long flags;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2006-12-10 10:19:24 +00:00
|
|
|
if (!mapping)
|
|
|
|
return 1;
|
|
|
|
|
2014-02-06 20:04:24 +00:00
|
|
|
spin_lock_irqsave(&mapping->tree_lock, flags);
|
2006-12-10 10:19:24 +00:00
|
|
|
mapping2 = page_mapping(page);
|
|
|
|
if (mapping2) { /* Race with truncate? */
|
|
|
|
BUG_ON(mapping2 != mapping);
|
2007-07-17 11:03:34 +00:00
|
|
|
WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
|
2009-03-31 22:19:39 +00:00
|
|
|
account_page_dirtied(page, mapping);
|
2006-12-10 10:19:24 +00:00
|
|
|
radix_tree_tag_set(&mapping->page_tree,
|
|
|
|
page_index(page), PAGECACHE_TAG_DIRTY);
|
|
|
|
}
|
2014-02-06 20:04:24 +00:00
|
|
|
spin_unlock_irqrestore(&mapping->tree_lock, flags);
|
2006-12-10 10:19:24 +00:00
|
|
|
if (mapping->host) {
|
|
|
|
/* !PageAnon && !swapper_space */
|
|
|
|
__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2006-03-24 11:18:11 +00:00
|
|
|
return 1;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2006-03-24 11:18:11 +00:00
|
|
|
return 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(__set_page_dirty_nobuffers);
|
|
|
|
|
2011-08-08 21:22:00 +00:00
|
|
|
/*
|
|
|
|
* Call this whenever redirtying a page, to de-account the dirty counters
|
|
|
|
* (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
|
|
|
|
* counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
|
|
|
|
* systematic errors in balanced_dirty_ratelimit and the dirty pages position
|
|
|
|
* control.
|
|
|
|
*/
|
|
|
|
void account_page_redirty(struct page *page)
|
|
|
|
{
|
|
|
|
struct address_space *mapping = page->mapping;
|
|
|
|
if (mapping && mapping_cap_account_dirty(mapping)) {
|
|
|
|
current->nr_dirtied--;
|
|
|
|
dec_zone_page_state(page, NR_DIRTIED);
|
|
|
|
dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(account_page_redirty);
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* When a writepage implementation decides that it doesn't want to write this
|
|
|
|
* page for some reason, it should redirty the locked page via
|
|
|
|
* redirty_page_for_writepage() and it should then unlock the page and return 0
|
|
|
|
*/
|
|
|
|
int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
|
|
|
|
{
|
|
|
|
wbc->pages_skipped++;
|
2011-08-08 21:22:00 +00:00
|
|
|
account_page_redirty(page);
|
2005-04-16 22:20:36 +00:00
|
|
|
return __set_page_dirty_nobuffers(page);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(redirty_page_for_writepage);
|
|
|
|
|
|
|
|
/*
|
2009-09-16 09:50:14 +00:00
|
|
|
* Dirty a page.
|
|
|
|
*
|
|
|
|
* For pages with a mapping this should be done under the page lock
|
|
|
|
* for the benefit of asynchronous memory errors who prefer a consistent
|
|
|
|
* dirty state. This rule can be broken in some special cases,
|
|
|
|
* but should be better not to.
|
|
|
|
*
|
2005-04-16 22:20:36 +00:00
|
|
|
* If the mapping doesn't provide a set_page_dirty a_op, then
|
|
|
|
* just fall through and assume that it wants buffer_heads.
|
|
|
|
*/
|
2009-02-18 22:48:18 +00:00
|
|
|
int set_page_dirty(struct page *page)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
|
|
|
struct address_space *mapping = page_mapping(page);
|
|
|
|
|
|
|
|
if (likely(mapping)) {
|
|
|
|
int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
|
mm: reclaim invalidated page ASAP
invalidate_mapping_pages is very big hint to reclaimer. It means user
doesn't want to use the page any more. So in order to prevent working set
page eviction, this patch move the page into tail of inactive list by
PG_reclaim.
Please, remember that pages in inactive list are working set as well as
active list. If we don't move pages into inactive list's tail, pages near
by tail of inactive list can be evicted although we have a big clue about
useless pages. It's totally bad.
Now PG_readahead/PG_reclaim is shared. fe3cba17 added ClearPageReclaim
into clear_page_dirty_for_io for preventing fast reclaiming readahead
marker page.
In this series, PG_reclaim is used by invalidated page, too. If VM find
the page is invalidated and it's dirty, it sets PG_reclaim to reclaim
asap. Then, when the dirty page will be writeback,
clear_page_dirty_for_io will clear PG_reclaim unconditionally. It
disturbs this serie's goal.
I think it's okay to clear PG_readahead when the page is dirty, not
writeback time. So this patch moves ClearPageReadahead. In v4,
ClearPageReadahead in set_page_dirty has a problem which is reported by
Steven Barrett. It's due to compound page. Some driver(ex, audio) calls
set_page_dirty with compound page which isn't on LRU. but my patch does
ClearPageRelcaim on compound page. In non-CONFIG_PAGEFLAGS_EXTENDED, it
breaks PageTail flag.
I think it doesn't affect THP and pass my test with THP enabling but Cced
Andrea for double check.
Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Reported-by: Steven Barrett <damentz@liquorix.net>
Reviewed-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-22 23:32:54 +00:00
|
|
|
/*
|
|
|
|
* readahead/lru_deactivate_page could remain
|
|
|
|
* PG_readahead/PG_reclaim due to race with end_page_writeback
|
|
|
|
* About readahead, if the page is written, the flags would be
|
|
|
|
* reset. So no problem.
|
|
|
|
* About lru_deactivate_page, if the page is redirty, the flag
|
|
|
|
* will be reset. So no problem. but if the page is used by readahead
|
|
|
|
* it will confuse readahead and make it restart the size rampup
|
|
|
|
* process. But it's a trivial problem.
|
|
|
|
*/
|
|
|
|
ClearPageReclaim(page);
|
[PATCH] BLOCK: Make it possible to disable the block layer [try #6]
Make it possible to disable the block layer. Not all embedded devices require
it, some can make do with just JFFS2, NFS, ramfs, etc - none of which require
the block layer to be present.
This patch does the following:
(*) Introduces CONFIG_BLOCK to disable the block layer, buffering and blockdev
support.
(*) Adds dependencies on CONFIG_BLOCK to any configuration item that controls
an item that uses the block layer. This includes:
(*) Block I/O tracing.
(*) Disk partition code.
(*) All filesystems that are block based, eg: Ext3, ReiserFS, ISOFS.
(*) The SCSI layer. As far as I can tell, even SCSI chardevs use the
block layer to do scheduling. Some drivers that use SCSI facilities -
such as USB storage - end up disabled indirectly from this.
(*) Various block-based device drivers, such as IDE and the old CDROM
drivers.
(*) MTD blockdev handling and FTL.
(*) JFFS - which uses set_bdev_super(), something it could avoid doing by
taking a leaf out of JFFS2's book.
(*) Makes most of the contents of linux/blkdev.h, linux/buffer_head.h and
linux/elevator.h contingent on CONFIG_BLOCK being set. sector_div() is,
however, still used in places, and so is still available.
(*) Also made contingent are the contents of linux/mpage.h, linux/genhd.h and
parts of linux/fs.h.
(*) Makes a number of files in fs/ contingent on CONFIG_BLOCK.
(*) Makes mm/bounce.c (bounce buffering) contingent on CONFIG_BLOCK.
(*) set_page_dirty() doesn't call __set_page_dirty_buffers() if CONFIG_BLOCK
is not enabled.
(*) fs/no-block.c is created to hold out-of-line stubs and things that are
required when CONFIG_BLOCK is not set:
(*) Default blockdev file operations (to give error ENODEV on opening).
(*) Makes some /proc changes:
(*) /proc/devices does not list any blockdevs.
(*) /proc/diskstats and /proc/partitions are contingent on CONFIG_BLOCK.
(*) Makes some compat ioctl handling contingent on CONFIG_BLOCK.
(*) If CONFIG_BLOCK is not defined, makes sys_quotactl() return -ENODEV if
given command other than Q_SYNC or if a special device is specified.
(*) In init/do_mounts.c, no reference is made to the blockdev routines if
CONFIG_BLOCK is not defined. This does not prohibit NFS roots or JFFS2.
(*) The bdflush, ioprio_set and ioprio_get syscalls can now be absent (return
error ENOSYS by way of cond_syscall if so).
(*) The seclvl_bd_claim() and seclvl_bd_release() security calls do nothing if
CONFIG_BLOCK is not set, since they can't then happen.
Signed-Off-By: David Howells <dhowells@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2006-09-30 18:45:40 +00:00
|
|
|
#ifdef CONFIG_BLOCK
|
|
|
|
if (!spd)
|
|
|
|
spd = __set_page_dirty_buffers;
|
|
|
|
#endif
|
|
|
|
return (*spd)(page);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2006-03-24 11:18:11 +00:00
|
|
|
if (!PageDirty(page)) {
|
|
|
|
if (!TestSetPageDirty(page))
|
|
|
|
return 1;
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(set_page_dirty);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* set_page_dirty() is racy if the caller has no reference against
|
|
|
|
* page->mapping->host, and if the page is unlocked. This is because another
|
|
|
|
* CPU could truncate the page off the mapping and then free the mapping.
|
|
|
|
*
|
|
|
|
* Usually, the page _is_ locked, or the caller is a user-space process which
|
|
|
|
* holds a reference on the inode by having an open file.
|
|
|
|
*
|
|
|
|
* In other cases, the page should be locked before running set_page_dirty().
|
|
|
|
*/
|
|
|
|
int set_page_dirty_lock(struct page *page)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
2011-03-10 07:52:07 +00:00
|
|
|
lock_page(page);
|
2005-04-16 22:20:36 +00:00
|
|
|
ret = set_page_dirty(page);
|
|
|
|
unlock_page(page);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(set_page_dirty_lock);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Clear a page's dirty flag, while caring for dirty memory accounting.
|
|
|
|
* Returns true if the page was previously dirty.
|
|
|
|
*
|
|
|
|
* This is for preparing to put the page under writeout. We leave the page
|
|
|
|
* tagged as dirty in the radix tree so that a concurrent write-for-sync
|
|
|
|
* can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
|
|
|
|
* implementation will run either set_page_writeback() or set_page_dirty(),
|
|
|
|
* at which stage we bring the page's dirty flag and radix-tree dirty tag
|
|
|
|
* back into sync.
|
|
|
|
*
|
|
|
|
* This incoherency between the page's dirty flag and radix-tree tag is
|
|
|
|
* unfortunate, but it only exists while the page is locked.
|
|
|
|
*/
|
|
|
|
int clear_page_dirty_for_io(struct page *page)
|
|
|
|
{
|
|
|
|
struct address_space *mapping = page_mapping(page);
|
|
|
|
|
2007-07-19 08:47:22 +00:00
|
|
|
BUG_ON(!PageLocked(page));
|
|
|
|
|
VM: Fix nasty and subtle race in shared mmap'ed page writeback
The VM layer (on the face of it, fairly reasonably) expected that when
it does a ->writepage() call to the filesystem, it would write out the
full page at that point in time. Especially since it had earlier marked
the whole page dirty with "set_page_dirty()".
But that isn't actually the case: ->writepage() does not actually write
a page, it writes the parts of the page that have been explicitly marked
dirty before, *and* that had not got written out for other reasons since
the last time we told it they were dirty.
That last caveat is the important one.
Which _most_ of the time ends up being the whole page (since we had
called "set_page_dirty()" on the page earlier), but if the filesystem
had done any dirty flushing of its own (for example, to honor some
internal write ordering guarantees), it might end up doing only a
partial page IO (or none at all) when ->writepage() is actually called.
That is the correct thing in general (since we actually often _want_
only the known-dirty parts of the page to be written out), but the
shared dirty page handling had implicitly forgotten about these details,
and had a number of cases where it was doing just the "->writepage()"
part, without telling the low-level filesystem that the whole page might
have been re-dirtied as part of being mapped writably into user space.
Since most of the time the FS did actually write out the full page, we
didn't notice this for a loong time, and this needed some really odd
patterns to trigger. But it caused occasional corruption with rtorrent
and with the Debian "apt" database, because both use shared mmaps to
update the end result.
This fixes it. Finally. After way too much hair-pulling.
Acked-by: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: Martin J. Bligh <mbligh@google.com>
Acked-by: Martin Michlmayr <tbm@cyrius.com>
Acked-by: Martin Johansson <martin@fatbob.nu>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Andrei Popa <andrei.popa@i-neo.ro>
Cc: High Dickins <hugh@veritas.com>
Cc: Andrew Morton <akpm@osdl.org>,
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Segher Boessenkool <segher@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Gordon Farquharson <gordonfarquharson@gmail.com>
Cc: Guillaume Chazarain <guichaz@yahoo.fr>
Cc: Theodore Tso <tytso@mit.edu>
Cc: Kenneth Cheng <kenneth.w.chen@intel.com>
Cc: Tobias Diedrich <ranma@tdiedrich.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-29 18:00:58 +00:00
|
|
|
if (mapping && mapping_cap_account_dirty(mapping)) {
|
|
|
|
/*
|
|
|
|
* Yes, Virginia, this is indeed insane.
|
|
|
|
*
|
|
|
|
* We use this sequence to make sure that
|
|
|
|
* (a) we account for dirty stats properly
|
|
|
|
* (b) we tell the low-level filesystem to
|
|
|
|
* mark the whole page dirty if it was
|
|
|
|
* dirty in a pagetable. Only to then
|
|
|
|
* (c) clean the page again and return 1 to
|
|
|
|
* cause the writeback.
|
|
|
|
*
|
|
|
|
* This way we avoid all nasty races with the
|
|
|
|
* dirty bit in multiple places and clearing
|
|
|
|
* them concurrently from different threads.
|
|
|
|
*
|
|
|
|
* Note! Normally the "set_page_dirty(page)"
|
|
|
|
* has no effect on the actual dirty bit - since
|
|
|
|
* that will already usually be set. But we
|
|
|
|
* need the side effects, and it can help us
|
|
|
|
* avoid races.
|
|
|
|
*
|
|
|
|
* We basically use the page "master dirty bit"
|
|
|
|
* as a serialization point for all the different
|
|
|
|
* threads doing their things.
|
|
|
|
*/
|
|
|
|
if (page_mkclean(page))
|
|
|
|
set_page_dirty(page);
|
2007-07-19 08:47:22 +00:00
|
|
|
/*
|
|
|
|
* We carefully synchronise fault handlers against
|
|
|
|
* installing a dirty pte and marking the page dirty
|
|
|
|
* at this point. We do this by having them hold the
|
|
|
|
* page lock at some point after installing their
|
|
|
|
* pte, but before marking the page dirty.
|
|
|
|
* Pages are always locked coming in here, so we get
|
|
|
|
* the desired exclusion. See mm/memory.c:do_wp_page()
|
|
|
|
* for more comments.
|
|
|
|
*/
|
VM: Fix nasty and subtle race in shared mmap'ed page writeback
The VM layer (on the face of it, fairly reasonably) expected that when
it does a ->writepage() call to the filesystem, it would write out the
full page at that point in time. Especially since it had earlier marked
the whole page dirty with "set_page_dirty()".
But that isn't actually the case: ->writepage() does not actually write
a page, it writes the parts of the page that have been explicitly marked
dirty before, *and* that had not got written out for other reasons since
the last time we told it they were dirty.
That last caveat is the important one.
Which _most_ of the time ends up being the whole page (since we had
called "set_page_dirty()" on the page earlier), but if the filesystem
had done any dirty flushing of its own (for example, to honor some
internal write ordering guarantees), it might end up doing only a
partial page IO (or none at all) when ->writepage() is actually called.
That is the correct thing in general (since we actually often _want_
only the known-dirty parts of the page to be written out), but the
shared dirty page handling had implicitly forgotten about these details,
and had a number of cases where it was doing just the "->writepage()"
part, without telling the low-level filesystem that the whole page might
have been re-dirtied as part of being mapped writably into user space.
Since most of the time the FS did actually write out the full page, we
didn't notice this for a loong time, and this needed some really odd
patterns to trigger. But it caused occasional corruption with rtorrent
and with the Debian "apt" database, because both use shared mmaps to
update the end result.
This fixes it. Finally. After way too much hair-pulling.
Acked-by: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: Martin J. Bligh <mbligh@google.com>
Acked-by: Martin Michlmayr <tbm@cyrius.com>
Acked-by: Martin Johansson <martin@fatbob.nu>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Andrei Popa <andrei.popa@i-neo.ro>
Cc: High Dickins <hugh@veritas.com>
Cc: Andrew Morton <akpm@osdl.org>,
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Segher Boessenkool <segher@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Gordon Farquharson <gordonfarquharson@gmail.com>
Cc: Guillaume Chazarain <guichaz@yahoo.fr>
Cc: Theodore Tso <tytso@mit.edu>
Cc: Kenneth Cheng <kenneth.w.chen@intel.com>
Cc: Tobias Diedrich <ranma@tdiedrich.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-29 18:00:58 +00:00
|
|
|
if (TestClearPageDirty(page)) {
|
2006-12-10 10:19:24 +00:00
|
|
|
dec_zone_page_state(page, NR_FILE_DIRTY);
|
2007-10-17 06:25:47 +00:00
|
|
|
dec_bdi_stat(mapping->backing_dev_info,
|
|
|
|
BDI_RECLAIMABLE);
|
VM: Fix nasty and subtle race in shared mmap'ed page writeback
The VM layer (on the face of it, fairly reasonably) expected that when
it does a ->writepage() call to the filesystem, it would write out the
full page at that point in time. Especially since it had earlier marked
the whole page dirty with "set_page_dirty()".
But that isn't actually the case: ->writepage() does not actually write
a page, it writes the parts of the page that have been explicitly marked
dirty before, *and* that had not got written out for other reasons since
the last time we told it they were dirty.
That last caveat is the important one.
Which _most_ of the time ends up being the whole page (since we had
called "set_page_dirty()" on the page earlier), but if the filesystem
had done any dirty flushing of its own (for example, to honor some
internal write ordering guarantees), it might end up doing only a
partial page IO (or none at all) when ->writepage() is actually called.
That is the correct thing in general (since we actually often _want_
only the known-dirty parts of the page to be written out), but the
shared dirty page handling had implicitly forgotten about these details,
and had a number of cases where it was doing just the "->writepage()"
part, without telling the low-level filesystem that the whole page might
have been re-dirtied as part of being mapped writably into user space.
Since most of the time the FS did actually write out the full page, we
didn't notice this for a loong time, and this needed some really odd
patterns to trigger. But it caused occasional corruption with rtorrent
and with the Debian "apt" database, because both use shared mmaps to
update the end result.
This fixes it. Finally. After way too much hair-pulling.
Acked-by: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: Martin J. Bligh <mbligh@google.com>
Acked-by: Martin Michlmayr <tbm@cyrius.com>
Acked-by: Martin Johansson <martin@fatbob.nu>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Andrei Popa <andrei.popa@i-neo.ro>
Cc: High Dickins <hugh@veritas.com>
Cc: Andrew Morton <akpm@osdl.org>,
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Segher Boessenkool <segher@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Gordon Farquharson <gordonfarquharson@gmail.com>
Cc: Guillaume Chazarain <guichaz@yahoo.fr>
Cc: Theodore Tso <tytso@mit.edu>
Cc: Kenneth Cheng <kenneth.w.chen@intel.com>
Cc: Tobias Diedrich <ranma@tdiedrich.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-29 18:00:58 +00:00
|
|
|
return 1;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
VM: Fix nasty and subtle race in shared mmap'ed page writeback
The VM layer (on the face of it, fairly reasonably) expected that when
it does a ->writepage() call to the filesystem, it would write out the
full page at that point in time. Especially since it had earlier marked
the whole page dirty with "set_page_dirty()".
But that isn't actually the case: ->writepage() does not actually write
a page, it writes the parts of the page that have been explicitly marked
dirty before, *and* that had not got written out for other reasons since
the last time we told it they were dirty.
That last caveat is the important one.
Which _most_ of the time ends up being the whole page (since we had
called "set_page_dirty()" on the page earlier), but if the filesystem
had done any dirty flushing of its own (for example, to honor some
internal write ordering guarantees), it might end up doing only a
partial page IO (or none at all) when ->writepage() is actually called.
That is the correct thing in general (since we actually often _want_
only the known-dirty parts of the page to be written out), but the
shared dirty page handling had implicitly forgotten about these details,
and had a number of cases where it was doing just the "->writepage()"
part, without telling the low-level filesystem that the whole page might
have been re-dirtied as part of being mapped writably into user space.
Since most of the time the FS did actually write out the full page, we
didn't notice this for a loong time, and this needed some really odd
patterns to trigger. But it caused occasional corruption with rtorrent
and with the Debian "apt" database, because both use shared mmaps to
update the end result.
This fixes it. Finally. After way too much hair-pulling.
Acked-by: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: Martin J. Bligh <mbligh@google.com>
Acked-by: Martin Michlmayr <tbm@cyrius.com>
Acked-by: Martin Johansson <martin@fatbob.nu>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Andrei Popa <andrei.popa@i-neo.ro>
Cc: High Dickins <hugh@veritas.com>
Cc: Andrew Morton <akpm@osdl.org>,
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Segher Boessenkool <segher@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Gordon Farquharson <gordonfarquharson@gmail.com>
Cc: Guillaume Chazarain <guichaz@yahoo.fr>
Cc: Theodore Tso <tytso@mit.edu>
Cc: Kenneth Cheng <kenneth.w.chen@intel.com>
Cc: Tobias Diedrich <ranma@tdiedrich.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-29 18:00:58 +00:00
|
|
|
return 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
VM: Fix nasty and subtle race in shared mmap'ed page writeback
The VM layer (on the face of it, fairly reasonably) expected that when
it does a ->writepage() call to the filesystem, it would write out the
full page at that point in time. Especially since it had earlier marked
the whole page dirty with "set_page_dirty()".
But that isn't actually the case: ->writepage() does not actually write
a page, it writes the parts of the page that have been explicitly marked
dirty before, *and* that had not got written out for other reasons since
the last time we told it they were dirty.
That last caveat is the important one.
Which _most_ of the time ends up being the whole page (since we had
called "set_page_dirty()" on the page earlier), but if the filesystem
had done any dirty flushing of its own (for example, to honor some
internal write ordering guarantees), it might end up doing only a
partial page IO (or none at all) when ->writepage() is actually called.
That is the correct thing in general (since we actually often _want_
only the known-dirty parts of the page to be written out), but the
shared dirty page handling had implicitly forgotten about these details,
and had a number of cases where it was doing just the "->writepage()"
part, without telling the low-level filesystem that the whole page might
have been re-dirtied as part of being mapped writably into user space.
Since most of the time the FS did actually write out the full page, we
didn't notice this for a loong time, and this needed some really odd
patterns to trigger. But it caused occasional corruption with rtorrent
and with the Debian "apt" database, because both use shared mmaps to
update the end result.
This fixes it. Finally. After way too much hair-pulling.
Acked-by: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: Martin J. Bligh <mbligh@google.com>
Acked-by: Martin Michlmayr <tbm@cyrius.com>
Acked-by: Martin Johansson <martin@fatbob.nu>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Andrei Popa <andrei.popa@i-neo.ro>
Cc: High Dickins <hugh@veritas.com>
Cc: Andrew Morton <akpm@osdl.org>,
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Segher Boessenkool <segher@kernel.crashing.org>
Cc: David Miller <davem@davemloft.net>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Gordon Farquharson <gordonfarquharson@gmail.com>
Cc: Guillaume Chazarain <guichaz@yahoo.fr>
Cc: Theodore Tso <tytso@mit.edu>
Cc: Kenneth Cheng <kenneth.w.chen@intel.com>
Cc: Tobias Diedrich <ranma@tdiedrich.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-29 18:00:58 +00:00
|
|
|
return TestClearPageDirty(page);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2005-11-18 09:10:53 +00:00
|
|
|
EXPORT_SYMBOL(clear_page_dirty_for_io);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
int test_clear_page_writeback(struct page *page)
|
|
|
|
{
|
|
|
|
struct address_space *mapping = page_mapping(page);
|
|
|
|
int ret;
|
2013-09-12 22:13:53 +00:00
|
|
|
bool locked;
|
|
|
|
unsigned long memcg_flags;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2013-09-12 22:13:53 +00:00
|
|
|
mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
if (mapping) {
|
2007-10-17 06:25:48 +00:00
|
|
|
struct backing_dev_info *bdi = mapping->backing_dev_info;
|
2005-04-16 22:20:36 +00:00
|
|
|
unsigned long flags;
|
|
|
|
|
2008-07-26 02:45:32 +00:00
|
|
|
spin_lock_irqsave(&mapping->tree_lock, flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
ret = TestClearPageWriteback(page);
|
2007-10-17 06:25:48 +00:00
|
|
|
if (ret) {
|
2005-04-16 22:20:36 +00:00
|
|
|
radix_tree_tag_clear(&mapping->page_tree,
|
|
|
|
page_index(page),
|
|
|
|
PAGECACHE_TAG_WRITEBACK);
|
2008-04-30 07:54:37 +00:00
|
|
|
if (bdi_cap_account_writeback(bdi)) {
|
2007-10-17 06:25:48 +00:00
|
|
|
__dec_bdi_stat(bdi, BDI_WRITEBACK);
|
2007-10-17 06:25:50 +00:00
|
|
|
__bdi_writeout_inc(bdi);
|
|
|
|
}
|
2007-10-17 06:25:48 +00:00
|
|
|
}
|
2008-07-26 02:45:32 +00:00
|
|
|
spin_unlock_irqrestore(&mapping->tree_lock, flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
} else {
|
|
|
|
ret = TestClearPageWriteback(page);
|
|
|
|
}
|
2011-07-26 00:12:37 +00:00
|
|
|
if (ret) {
|
2013-09-12 22:13:53 +00:00
|
|
|
mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
|
2007-07-19 08:49:17 +00:00
|
|
|
dec_zone_page_state(page, NR_WRITEBACK);
|
2011-07-26 00:12:37 +00:00
|
|
|
inc_zone_page_state(page, NR_WRITTEN);
|
|
|
|
}
|
2013-09-12 22:13:53 +00:00
|
|
|
mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2014-05-12 12:12:25 +00:00
|
|
|
int __test_set_page_writeback(struct page *page, bool keep_write)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
|
|
|
struct address_space *mapping = page_mapping(page);
|
|
|
|
int ret;
|
2013-09-12 22:13:53 +00:00
|
|
|
bool locked;
|
|
|
|
unsigned long memcg_flags;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2013-09-12 22:13:53 +00:00
|
|
|
mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
if (mapping) {
|
2007-10-17 06:25:48 +00:00
|
|
|
struct backing_dev_info *bdi = mapping->backing_dev_info;
|
2005-04-16 22:20:36 +00:00
|
|
|
unsigned long flags;
|
|
|
|
|
2008-07-26 02:45:32 +00:00
|
|
|
spin_lock_irqsave(&mapping->tree_lock, flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
ret = TestSetPageWriteback(page);
|
2007-10-17 06:25:48 +00:00
|
|
|
if (!ret) {
|
2005-04-16 22:20:36 +00:00
|
|
|
radix_tree_tag_set(&mapping->page_tree,
|
|
|
|
page_index(page),
|
|
|
|
PAGECACHE_TAG_WRITEBACK);
|
2008-04-30 07:54:37 +00:00
|
|
|
if (bdi_cap_account_writeback(bdi))
|
2007-10-17 06:25:48 +00:00
|
|
|
__inc_bdi_stat(bdi, BDI_WRITEBACK);
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
if (!PageDirty(page))
|
|
|
|
radix_tree_tag_clear(&mapping->page_tree,
|
|
|
|
page_index(page),
|
|
|
|
PAGECACHE_TAG_DIRTY);
|
2014-05-12 12:12:25 +00:00
|
|
|
if (!keep_write)
|
|
|
|
radix_tree_tag_clear(&mapping->page_tree,
|
|
|
|
page_index(page),
|
|
|
|
PAGECACHE_TAG_TOWRITE);
|
2008-07-26 02:45:32 +00:00
|
|
|
spin_unlock_irqrestore(&mapping->tree_lock, flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
} else {
|
|
|
|
ret = TestSetPageWriteback(page);
|
|
|
|
}
|
2007-07-19 08:49:17 +00:00
|
|
|
if (!ret)
|
2010-10-26 21:21:33 +00:00
|
|
|
account_page_writeback(page);
|
2013-09-12 22:13:53 +00:00
|
|
|
mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
return ret;
|
|
|
|
|
|
|
|
}
|
2014-05-12 12:12:25 +00:00
|
|
|
EXPORT_SYMBOL(__test_set_page_writeback);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
/*
|
2007-10-16 08:24:40 +00:00
|
|
|
* Return true if any of the pages in the mapping are marked with the
|
2005-04-16 22:20:36 +00:00
|
|
|
* passed tag.
|
|
|
|
*/
|
|
|
|
int mapping_tagged(struct address_space *mapping, int tag)
|
|
|
|
{
|
2011-07-26 00:12:31 +00:00
|
|
|
return radix_tree_tagged(&mapping->page_tree, tag);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(mapping_tagged);
|
mm: only enforce stable page writes if the backing device requires it
Create a helper function to check if a backing device requires stable
page writes and, if so, performs the necessary wait. Then, make it so
that all points in the memory manager that handle making pages writable
use the helper function. This should provide stable page write support
to most filesystems, while eliminating unnecessary waiting for devices
that don't require the feature.
Before this patchset, all filesystems would block, regardless of whether
or not it was necessary. ext3 would wait, but still generate occasional
checksum errors. The network filesystems were left to do their own
thing, so they'd wait too.
After this patchset, all the disk filesystems except ext3 and btrfs will
wait only if the hardware requires it. ext3 (if necessary) snapshots
pages instead of blocking, and btrfs provides its own bdi so the mm will
never wait. Network filesystems haven't been touched, so either they
provide their own stable page guarantees or they don't block at all.
The blocking behavior is back to what it was before 3.0 if you don't
have a disk requiring stable page writes.
Here's the result of using dbench to test latency on ext2:
3.8.0-rc3:
Operation Count AvgLat MaxLat
----------------------------------------
WriteX 109347 0.028 59.817
ReadX 347180 0.004 3.391
Flush 15514 29.828 287.283
Throughput 57.429 MB/sec 4 clients 4 procs max_latency=287.290 ms
3.8.0-rc3 + patches:
WriteX 105556 0.029 4.273
ReadX 335004 0.005 4.112
Flush 14982 30.540 298.634
Throughput 55.4496 MB/sec 4 clients 4 procs max_latency=298.650 ms
As you can see, the maximum write latency drops considerably with this
patch enabled. The other filesystems (ext3/ext4/xfs/btrfs) behave
similarly, but see the cover letter for those results.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Acked-by: Steven Whitehouse <swhiteho@redhat.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Artem Bityutskiy <dedekind1@gmail.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Mark Fasheh <mfasheh@suse.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Eric Van Hensbergen <ericvh@gmail.com>
Cc: Ron Minnich <rminnich@sandia.gov>
Cc: Latchesar Ionkov <lucho@ionkov.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-22 00:42:51 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* wait_for_stable_page() - wait for writeback to finish, if necessary.
|
|
|
|
* @page: The page to wait on.
|
|
|
|
*
|
|
|
|
* This function determines if the given page is related to a backing device
|
|
|
|
* that requires page contents to be held stable during writeback. If so, then
|
|
|
|
* it will wait for any pending writeback to complete.
|
|
|
|
*/
|
|
|
|
void wait_for_stable_page(struct page *page)
|
|
|
|
{
|
|
|
|
struct address_space *mapping = page_mapping(page);
|
|
|
|
struct backing_dev_info *bdi = mapping->backing_dev_info;
|
|
|
|
|
|
|
|
if (!bdi_cap_stable_pages_required(bdi))
|
|
|
|
return;
|
|
|
|
|
|
|
|
wait_on_page_writeback(page);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(wait_for_stable_page);
|