Skip to content
  1. Jun 08, 2011
    • Christoph Hellwig's avatar
      writeback: split inode_wb_list_lock into bdi_writeback.list_lock · f758eeab
      Christoph Hellwig authored 
      

Split the global inode_wb_list_lock into a per-bdi_writeback list_lock,
as it's currently the most contended lock in the system for metadata
heavy workloads.  It won't help for single-filesystem workloads for
which we'll need the I/O-less balance_dirty_pages, but at least we
can dedicate a cpu to spinning on each bdi now for larger systems.

Based on earlier patches from Nick Piggin and Dave Chinner.

It reduces lock contentions to 1/4 in this test case:
10 HDD JBOD, 100 dd on each disk, XFS, 6GB ram

lock_stat version 0.3
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
                              class name    con-bounces    contentions   waittime-min   waittime-max waittime-total    acq-bounces   acquisitions   holdtime-min   holdtime-max holdtime-total
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
vanilla 2.6.39-rc3:
                      inode_wb_list_lock:         42590          44433           0.12         147.74      144127.35         252274         886792           0.08         121.34      917211.23
                      ------------------
                      inode_wb_list_lock              2          [<ffffffff81165da5>] bdev_inode_switch_bdi+0x29/0x85
                      inode_wb_list_lock             34          [<ffffffff8115bd0b>] inode_wb_list_del+0x22/0x49
                      inode_wb_list_lock          12893          [<ffffffff8115bb53>] __mark_inode_dirty+0x170/0x1d0
                      inode_wb_list_lock          10702          [<ffffffff8115afef>] writeback_single_inode+0x16d/0x20a
                      ------------------
                      inode_wb_list_lock              2          [<ffffffff81165da5>] bdev_inode_switch_bdi+0x29/0x85
                      inode_wb_list_lock             19          [<ffffffff8115bd0b>] inode_wb_list_del+0x22/0x49
                      inode_wb_list_lock           5550          [<ffffffff8115bb53>] __mark_inode_dirty+0x170/0x1d0
                      inode_wb_list_lock           8511          [<ffffffff8115b4ad>] writeback_sb_inodes+0x10f/0x157

2.6.39-rc3 + patch:
                &(&wb->list_lock)->rlock:         11383          11657           0.14         151.69       40429.51          90825         527918           0.11         145.90      556843.37
                ------------------------
                &(&wb->list_lock)->rlock             10          [<ffffffff8115b189>] inode_wb_list_del+0x5f/0x86
                &(&wb->list_lock)->rlock           1493          [<ffffffff8115b1ed>] writeback_inodes_wb+0x3d/0x150
                &(&wb->list_lock)->rlock           3652          [<ffffffff8115a8e9>] writeback_sb_inodes+0x123/0x16f
                &(&wb->list_lock)->rlock           1412          [<ffffffff8115a38e>] writeback_single_inode+0x17f/0x223
                ------------------------
                &(&wb->list_lock)->rlock              3          [<ffffffff8110b5af>] bdi_lock_two+0x46/0x4b
                &(&wb->list_lock)->rlock              6          [<ffffffff8115b189>] inode_wb_list_del+0x5f/0x86
                &(&wb->list_lock)->rlock           2061          [<ffffffff8115af97>] __mark_inode_dirty+0x173/0x1cf
                &(&wb->list_lock)->rlock           2629          [<ffffffff8115a8e9>] writeback_sb_inodes+0x123/0x16f

hughd@google.com: fix recursive lock when bdi_lock_two() is called with new the same as old
akpm@linux-foundation.org: cleanup bdev_inode_switch_bdi() comment

Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
Signed-off-by: default avatarHugh Dickins <hughd@google.com>
Signed-off-by: default avatarAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: default avatarWu Fengguang <fengguang.wu@intel.com>
      f758eeab
    • Wu Fengguang's avatar
      writeback: refill b_io iff empty · 424b351f
      Wu Fengguang authored 
      

There is no point to carry different refill policies between for_kupdate
and other type of works. Use a consistent "refill b_io iff empty" policy
which can guarantee fairness in an easy to understand way.

A b_io refill will setup a _fixed_ work set with all currently eligible
inodes and start a new round of walk through b_io. The "fixed" work set
means no new inodes will be added to the work set during the walk.
Only when a complete walk over b_io is done, new inodes that are
eligible at the time will be enqueued and the walk be started over.

This procedure provides fairness among the inodes because it guarantees
each inode to be synced once and only once at each round. So all inodes
will be free from starvations.

This change relies on wb_writeback() to keep retrying as long as we made
some progress on cleaning some pages and/or inodes. Without that ability,
the old logic on background works relies on aggressively queuing all
eligible inodes into b_io at every time. But that's not a guarantee.

The below test script completes a slightly faster now:

             2.6.39-rc3	  2.6.39-rc3-dyn-expire+
------------------------------------------------
all elapsed     256.043      252.367
stddev           24.381       12.530

tar elapsed      30.097       28.808
dd  elapsed      13.214       11.782

	#!/bin/zsh

	cp /c/linux-2.6.38.3.tar.bz2 /dev/shm/

	umount /dev/sda7
	mkfs.xfs -f /dev/sda7
	mount /dev/sda7 /fs

	echo 3 > /proc/sys/vm/drop_caches

	tic=$(cat /proc/uptime|cut -d' ' -f2)

	cd /fs
	time tar jxf /dev/shm/linux-2.6.38.3.tar.bz2 &
	time dd if=/dev/zero of=/fs/zero bs=1M count=1000 &

	wait
	sync
	tac=$(cat /proc/uptime|cut -d' ' -f2)
	echo elapsed: $((tac - tic))

It maintains roughly the same small vs. large file writeout shares, and
offers large files better chances to be written in nice 4M chunks.

Analyzes from Dave Chinner in great details:

Let's say we have lots of inodes with 100 dirty pages being created,
and one large writeback going on. We expire 8 new inodes for every
1024 pages we write back.

With the old code, we do:

	b_more_io (large inode) -> b_io (1l)
	8 newly expired inodes -> b_io (1l, 8s)

	writeback  large inode 1024 pages -> b_more_io

	b_more_io (large inode) -> b_io (8s, 1l)
	8 newly expired inodes -> b_io (8s, 1l, 8s)

	writeback  8 small inodes 800 pages
		   1 large inode 224 pages -> b_more_io

	b_more_io (large inode) -> b_io (8s, 1l)
	8 newly expired inodes -> b_io (8s, 1l, 8s)
	.....

Your new code:

	b_more_io (large inode) -> b_io (1l)
	8 newly expired inodes -> b_io (1l, 8s)

	writeback  large inode 1024 pages -> b_more_io
	(b_io == 8s)
	writeback  8 small inodes 800 pages

	b_io empty: (1800 pages written)
		b_more_io (large inode) -> b_io (1l)
		14 newly expired inodes -> b_io (1l, 14s)

	writeback  large inode 1024 pages -> b_more_io
	(b_io == 14s)
	writeback  10 small inodes 1000 pages
		   1 small inode 24 pages -> b_more_io (1l, 1s(24))
	writeback  5 small inodes 500 pages
	b_io empty: (2548 pages written)
		b_more_io (large inode) -> b_io (1l, 1s(24))
		20 newly expired inodes -> b_io (1l, 1s(24), 20s)
	......

Rough progression of pages written at b_io refill:

Old code:

	total	large file	% of writeback
	1024	224		21.9% (fixed)

New code:
	total	large file	% of writeback
	1800	1024		~55%
	2550	1024		~40%
	3050	1024		~33%
	3500	1024		~29%
	3950	1024		~26%
	4250	1024		~24%
	4500	1024		~22.7%
	4700	1024		~21.7%
	4800	1024		~21.3%
	4800	1024		~21.3%
	(pretty much steady state from here)

Ok, so the steady state is reached with a similar percentage of
writeback to the large file as the existing code. Ok, that's good,
but providing some evidence that is doesn't change the shared of
writeback to the large should be in the commit message ;)

The other advantage to this is that we always write 1024 page chunks
to the large file, rather than smaller "whatever remains" chunks.

CC: Jan Kara <jack@suse.cz>
Acked-by: default avatarMel Gorman <mel@csn.ul.ie>
Signed-off-by: default avatarWu Fengguang <fengguang.wu@intel.com>
      424b351f
    • Wu Fengguang's avatar
      writeback: the kupdate expire timestamp should be a moving target · ba9aa839
      Wu Fengguang authored 
      

Dynamically compute the dirty expire timestamp at queue_io() time.

writeback_control.older_than_this used to be determined at entrance to
the kupdate writeback work. This _static_ timestamp may go stale if the
kupdate work runs on and on. The flusher may then stuck with some old
busy inodes, never considering newly expired inodes thereafter.

This has two possible problems:

- It is unfair for a large dirty inode to delay (for a long time) the
  writeback of small dirty inodes.

- As time goes by, the large and busy dirty inode may contain only
  _freshly_ dirtied pages. Ignoring newly expired dirty inodes risks
  delaying the expired dirty pages to the end of LRU lists, triggering
  the evil pageout(). Nevertheless this patch merely addresses part
  of the problem.

v2: keep policy changes inside wb_writeback() and keep the
wbc.older_than_this visibility as suggested by Dave.

CC: Dave Chinner <david@fromorbit.com>
Acked-by: default avatarJan Kara <jack@suse.cz>
Acked-by: default avatarMel Gorman <mel@csn.ul.ie>
Signed-off-by: default avatarItaru Kitayama <kitayama@cl.bb4u.ne.jp>
Signed-off-by: default avatarWu Fengguang <fengguang.wu@intel.com>
      ba9aa839
    • Wu Fengguang's avatar
      writeback: try more writeback as long as something was written · e6fb6da2
      Wu Fengguang authored 
      

writeback_inodes_wb()/__writeback_inodes_sb() are not aggressive in that
they only populate possibly a subset of eligible inodes into b_io at
entrance time. When the queued set of inodes are all synced, they just
return, possibly with all queued inode pages written but still
wbc.nr_to_write > 0.

For kupdate and background writeback, there may be more eligible inodes
sitting in b_dirty when the current set of b_io inodes are completed. So
it is necessary to try another round of writeback as long as we made some
progress in this round. When there are no more eligible inodes, no more
inodes will be enqueued in queue_io(), hence nothing could/will be
synced and we may safely bail.

For example, imagine 100 inodes

        i0, i1, i2, ..., i90, i91, i99

At queue_io() time, i90-i99 happen to be expired and moved to s_io for
IO. When finished successfully, if their total size is less than
MAX_WRITEBACK_PAGES, nr_to_write will be > 0. Then wb_writeback() will
quit the background work (w/o this patch) while it's still over
background threshold. This will be a fairly normal/frequent case I guess.

Now that we do tagged sync and update inode->dirtied_when after the sync,
this change won't livelock sync(1).  I actually tried to write 1 page
per 1ms with this command

	write-and-fsync -n10000 -S 1000 -c 4096 /fs/test

and do sync(1) at the same time. The sync completes quickly on ext4,
xfs, btrfs.

Acked-by: default avatarJan Kara <jack@suse.cz>
Signed-off-by: default avatarWu Fengguang <fengguang.wu@intel.com>
      e6fb6da2
    • Wu Fengguang's avatar
      writeback: introduce writeback_control.inodes_written · cb9bd115
      Wu Fengguang authored 
      

The flusher works on dirty inodes in batches, and may quit prematurely
if the batch of inodes happen to be metadata-only dirtied: in this case
wbc->nr_to_write won't be decreased at all, which stands for "no pages
written" but also mis-interpreted as "no progress".

So introduce writeback_control.inodes_written to count the inodes get
cleaned from VFS POV.  A non-zero value means there are some progress on
writeback, in which case more writeback can be tried.

Acked-by: default avatarJan Kara <jack@suse.cz>
Acked-by: default avatarMel Gorman <mel@csn.ul.ie>
Signed-off-by: default avatarWu Fengguang <fengguang.wu@intel.com>
      cb9bd115
    • Wu Fengguang's avatar
      writeback: update dirtied_when for synced inode to prevent livelock · 94c3dcbb
      Wu Fengguang authored 
      Explicitly update .dirtied_when on synced inodes, so that they are no
longer considered for writeback in the next round.

It can prevent both of the following livelock schemes:

- while true; do echo data >> f; done
- while true; do touch f;        done (in theory)

The exact livelock condition is, during sync(1):

(1) no new inodes are dirtied
(2) an inode being actively dirtied

On (2), the inode will be tagged and synced with .nr_to_write=LONG_MAX.
When finished, it will be redirty_tail()ed because it's still dirty
and (.nr_to_write > 0). redirty_tail() won't update its ->dirtied_when
on condition (1). The sync work will then revisit it on the next
queue_io() and find it eligible again because its old ->dirtied_when
predates the sync work start time.

We'll do more aggressive "keep writeback as long as we wrote something"
logic in wb_writeback(). The "use LONG_MAX .nr_to_write" trick in commit
b9543dac

 ("writeback: avoid livelocking WB_SYNC_ALL writeback") will
no longer be enough to stop sync livelock.

Reviewed-by: default avatarJan Kara <jack@suse.cz>
Signed-off-by: default avatarWu Fengguang <fengguang.wu@intel.com>
      94c3dcbb
    • Wu Fengguang's avatar
      writeback: introduce .tagged_writepages for the WB_SYNC_NONE sync stage · 6e6938b6
      Wu Fengguang authored 
      sync(2) is performed in two stages: the WB_SYNC_NONE sync and the
WB_SYNC_ALL sync. Identify the first stage with .tagged_writepages and
do livelock prevention for it, too.

Jan's commit f446daae ("mm: implement writeback livelock avoidance
using page tagging") is a partial fix in that it only fixed the
WB_SYNC_ALL phase livelock.

Although ext4 is tested to no longer livelock with commit f446daae

,
it may due to some "redirty_tail() after pages_skipped" effect which
is by no means a guarantee for _all_ the file systems.

Note that writeback_inodes_sb() is called by not only sync(), they are
treated the same because the other callers also need livelock prevention.

Impact:  It changes the order in which pages/inodes are synced to disk.
Now in the WB_SYNC_NONE stage, it won't proceed to write the next inode
until finished with the current inode.

Acked-by: default avatarJan Kara <jack@suse.cz>
CC: Dave Chinner <david@fromorbit.com>
Signed-off-by: default avatarWu Fengguang <fengguang.wu@intel.com>
      6e6938b6
  3. Jun 06, 2011
  5. Jun 05, 2011
  7. Jun 04, 2011