LKML Archive on
help / color / mirror / Atom feed
From: Jan Kara <>
To: Andrew Morton <>
Subject: Re: [PATCH] Fix private_list handling
Date: Fri, 11 Jan 2008 15:21:31 +0100	[thread overview]
Message-ID: <> (raw)
In-Reply-To: <>

On Thu 10-01-08 16:36:35, Andrew Morton wrote:
> On Thu, 10 Jan 2008 16:55:13 +0100
> Jan Kara <> wrote:
> >   Hi,
> > 
> >   sorry for the previous empty email...
> > 
> >   Supriya noted in his testing that sometimes buffers removed by
> > __remove_assoc_queue() don't have b_assoc_mapping set (and thus IO error
> > won't be properly propagated). Actually, looking more into the code I found
> > there are some more races. The patch below should fix them. It survived
> > beating with LTP and fsstress on ext2 filesystem on my testing machine so
> > it should be reasonably bugfree... Andrew, would you put the patch into
> > -mm? Thanks.
> > 
> > 								Honza
> > -- 
> > Jan Kara <>
> > SUSE Labs, CR
> > ---
> > 
> > There are two possible races in handling of private_list in buffer cache.
> > 1) When fsync_buffers_list() processes a private_list, it clears
> > b_assoc_mapping and moves buffer to its private list. Now drop_buffers() comes,
> > sees a buffer is on list so it calls __remove_assoc_queue() which complains
> > about b_assoc_mapping being cleared (as it cannot propagate possible IO error).
> > This race has been actually observed in the wild.
> private_lock should prevent this race.
> Which call to drop_buffers() is the culprit?  The first one in
> try_to_free_buffers(), I assume?  The "can this still happen?" one?
> If so, it can happen.  How?  Perhaps this is a bug.
  Good question but I don't think so. The problem is that
fsync_buffers_list() drops the private_lock() e.g. when it does
wait_on_buffer(). And buffers on the list fsync_buffers_list() constructs
have b_assoc_mapping set to NULL but the test
!list_empty(bh->b_assoc_buffers) succeeds for them and thus
__remove_assoc_queue() is called and it complains.
  We could also silence the warning by leaving b_assoc_mapping set when we
move the buffer to the constructed list. But given the problem below
I've decided to do a more complete cleanup of the code.

> > 2) When fsync_buffers_list() processes a private_list,
> > mark_buffer_dirty_inode() can be called on bh which is already on the private
> > list of fsync_buffers_list(). As buffer is on some list (note that the check is
> > performed without private_lock), it is not readded to the mapping's
> > private_list and after fsync_buffers_list() finishes, we have a dirty buffer
> > which should be on private_list but it isn't. This race has not been reported,
> > probably because most (but not all) callers of mark_buffer_dirty_inode() hold
> > i_mutex and thus are serialized with fsync().
> Maybe fsync_buffers_list should put the buffer back onto private_list if it
> got dirtied again.
  Yes, that's what it does in my new version. Only the locking is a bit
subtle if we want to avoid taking private_lock in mark_buffer_dirty_inode()
when the buffer is already on some list...

Jan Kara <>

  reply	other threads:[~2008-01-11 14:21 UTC|newest]

Thread overview: 8+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2008-01-10 15:50 Jan Kara
2008-01-10 15:55 ` Jan Kara
2008-01-11  0:36   ` Andrew Morton
2008-01-11 14:21     ` Jan Kara [this message]
2008-01-11 23:33       ` Andrew Morton
2008-01-14 11:59         ` Jan Kara
2008-01-14 18:14         ` Jan Kara
2008-02-04 16:44 Jan Kara

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \ \ \ \ \ \
    --subject='Re: [PATCH] Fix private_list handling' \

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).