On Thu 11-07-13 13:58:32, Jan Kara wrote:
> On Thu 11-07-13 12:53:46, Jan Kara wrote:
> > On Wed 10-07-13 16:12:36, Paul Taysom wrote:
> > > The following commit introduced a 10x regression for
> > > syncing inodes in ext4 with relatime enabled where just
> > > the atime had been modified.
> > >
> > > commit 4ea425b63a3dfeb7707fc7cc7161c11a51e871ed
> > > Author: Jan Kara <jack@xxxxxxx>
> > > Date: Tue Jul 3 16:45:34 2012 +0200
> > > vfs: Avoid unnecessary WB_SYNC_NONE writeback during sys_sync and reorder sync passes
> > >
> > > See also: http://www.kernelhub.org/?msg=93100&p=2
> > >
> > > Fixed by putting back in the call to writeback_inodes_sb.
> > >
> > > I'll attach the test in a reply to this e-mail.
> > >
> > > The test starts by creating 512 files, syncing, reading one byte
> > > from each of those files, syncing, and then deleting each file
> > > and syncing. The time to do each sync is printed. The process
> > > is then repeated for 1024 files and then the next power of
> > > two up to 262144 files.
> > >
> > > Note, when running the test, the slow down doesn't always happen
> > > but most of the tests will show a slow down.
> > >
> > > In response to crbug.com/240422
> > >
> > > Signed-off-by: Paul Taysom <taysom@xxxxxxxxxxxx>
> > Thanks for report. Rather than blindly reverting the change, I'd like to
> > understand why you see so huge regression. As the changelog in the patch
> > says, flusher thread should be doing async writeback equivalent to the
> > removed one because it gets woken via wakeup_flusher_threads(). But my
> > guess is that for some reason we end up doing all the writeback from
> > sync_inodes_one_sb(). I'll try to reproduce your results and investigate...
> Hum, so it must be something timing sensitive. I wasn't able to reproduce
> the issue on my test machine in 4 runs of your test program. I was able to
> reproduce it on my laptop on every second run of the test program but once
> I've enabled some tracepoints, the issue disappeared and I didn't see it in
> about 10 runs.
>
> That being said I think that reverting my patch is just papering over the
> problem. We will do the async pass over inodes twice instead of once
> and thus the timing changes enough that you aren't able to observe the
> problem.
>
> I'm looking into this more...
So I finally understood what's going on. If the system has no dirty pages
at all wakeup_flusher_threads() will submit work with nr_pages == 0. So
wb_writeback() will bail out immediately without doing anything and all the
writeback is left for WB_SYNC_ALL pass of sync(1) which is slow. Attached
patch fixes the problem for me.
Honza
--
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR
>From 2e3d6f21ffa990780e9b25e11be31a6e0da13c79 Mon Sep 17 00:00:00 2001
From: Jan Kara <jack@xxxxxxx>
Date: Fri, 12 Jul 2013 17:30:07 +0200
Subject: [PATCH] writeback: Fix occasional slow sync(1)
In case when system contains no dirty pages, wakeup_flusher_threads()
will submit WB_SYNC_NONE writeback for 0 pages so wb_writeback() exits
immediately without doing anything. Thus sync(1) will write all the
dirty inodes from a WB_SYNC_ALL writeback pass which is slow.
Fix the problem by using get_nr_dirty_pages() in
wakeup_flusher_threads() instead of calculating number of dirty pages
manually. That function also takes number of dirty inodes into account.
CC: stable@xxxxxxxxxxxxxxx
Reported-by: Paul Taysom <taysom@xxxxxxxxxxxx>
Signed-off-by: Jan Kara <jack@xxxxxxx>
---
fs/fs-writeback.c | 6 ++----
1 file changed, 2 insertions(+), 4 deletions(-)
diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index a85ac4e..d0d70a8 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -1055,10 +1055,8 @@ void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
{
struct backing_dev_info *bdi;
- if (!nr_pages) {
- nr_pages = global_page_state(NR_FILE_DIRTY) +
- global_page_state(NR_UNSTABLE_NFS);
- }
+ if (!nr_pages)
+ nr_pages = get_nr_dirty_pages();
rcu_read_lock();
list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
--
1.8.1.4
