lkml.org 
[lkml]   [2010]   [Aug]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/5] writeback: stop periodic/background work on seeing sync works
On Tue, Aug 03, 2010 at 04:51:52AM +0800, Jan Kara wrote:
> On Fri 30-07-10 12:03:06, Wu Fengguang wrote:
> > On Fri, Jul 30, 2010 at 12:20:27AM +0800, Jan Kara wrote:
> > > On Thu 29-07-10 19:51:44, Wu Fengguang wrote:
> > > > The periodic/background writeback can run forever. So when any
> > > > sync work is enqueued, increase bdi->sync_works to notify the
> > > > active non-sync works to exit. Non-sync works queued after sync
> > > > works won't be affected.
> > > Hmm, wouldn't it be simpler logic to just make for_kupdate and
> > > for_background work always yield when there's some other work to do (as
> > > they are livelockable from the definition of the target they have) and
> > > make sure any other work isn't livelockable?
> >
> > Good idea!
> >
> > > The only downside is that
> > > non-livelockable work cannot be "fair" in the sense that we cannot switch
> > > inodes after writing MAX_WRITEBACK_PAGES.
> >
> > Cannot switch indoes _before_ finish with the current
> > MAX_WRITEBACK_PAGES batch?
> Well, even after writing all those MAX_WRITEBACK_PAGES. Because what you
> want to do in a non-livelockable work is: take inode, write it, never look at
> it again for this work. Because if you later return to the inode, it can
> have newer dirty pages and thus you cannot really avoid livelock. Of
> course, this all assumes .nr_to_write isn't set to something small. That
> avoids the livelock as well.

I do have a poor man's solution that can handle this case.
https://kerneltrap.org/mailarchive/linux-fsdevel/2009/10/7/6476473/thread
It may do more extra works, but will stop livelock in theory.

A related question is, what if some for_reclaim works get enqueued?
Shall we postpone the sync work as well? The global sync is not likely
to hit the dirty pages in a small memcg, or may take long time. It
seems not a high priority task though.

> > > I even had a patch for this but it's already outdated by now. But I
> > > can refresh it if we decide this is the way to go.
> >
> > I'm very interested in your old patch, would you post it? Let's see
> > which one is easier to work with :)
> OK, attached is the patch. I've rebased it against 2.6.35.
> Honza
> --
> Jan Kara <jack@suse.cz>
> SUSE Labs, CR

> From a6df0d4db148f983fe756df4791409db28dff459 Mon Sep 17 00:00:00 2001
> From: Jan Kara <jack@suse.cz>
> Date: Mon, 2 Aug 2010 22:30:25 +0200
> Subject: [PATCH] mm: Stop background writeback if there is other work queued for the thread
>
> Background writeback and kupdate-style writeback are easily livelockable
> (from a definition of their target). This is inconvenient because it can
> make sync(1) stall forever waiting on its queued work to be finished.
> Fix the problem by interrupting background and kupdate writeback if there
> is some other work to do. We can return to them after completing all the
> queued work.
>
> Signed-off-by: Jan Kara <jack@suse.cz>
> ---
> fs/fs-writeback.c | 8 ++++++++
> 1 files changed, 8 insertions(+), 0 deletions(-)
>
> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> index d5be169..542471e 100644
> --- a/fs/fs-writeback.c
> +++ b/fs/fs-writeback.c
> @@ -633,6 +633,14 @@ static long wb_writeback(struct bdi_writeback *wb,
> break;
>
> /*
> + * Background writeout and kupdate-style writeback are
> + * easily livelockable. Stop them if there is other work
> + * to do so that e.g. sync can proceed.
> + */
> + if ((work->for_background || work->for_kupdate) &&
> + !list_empty(&wb->bdi->work_list))
> + break;
> + /*

I like it. It's much simpler.

Reviewed-by: Wu Fengguang <fengguang.wu@intel.com>


Thanks,
Fengguang


\
 
 \ /
  Last update: 2010-08-03 05:03    [W:0.096 / U:0.700 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site