[PATCH v3] [RESEND] mm: Make snapshotting pages for stable writes a per-bio operation

[Jan Kara]

On Tue 09-04-13 14:04:32, Andrew Morton wrote:
> On Tue, 9 Apr 2013 11:06:17 -0700 "Darrick J. Wong" <darrick.wong at oracle.com> wrote:
> 
> > +		 * Here we write back pagecache data that may be mmaped. Since
> > +		 * we cannot afford to clean the page and set PageWriteback
> > +		 * here due to lock ordering (page lock ranks above transaction
> > +		 * start), the data can change while IO is in flight. Tell the
> > +		 * block layer it should bounce the bio pages if stable data
> > +		 * during write is required.
> 
> I think there are already ab/ba deadlocks between lock_page() and
> journal_start().  iirc one path was write(), I forget which was the
> other path.  This was 10+ years ago and nobody else noticed and I
> didn't know how to fix it so I didn't tell anyone ;)
  Hum, I don't think they are there anymore ;)

> It would be neat to be able to hook things like journal_start() into
> lockdep but I don't think that lockdep has easy provision for wiring
> oddball things into its mechanisms.
  Actually if you look at fs/jbd/transaction.c:start_this_handle(), you
will notice that it is wired into lockdep (the lock_map_acquire() call
there). And it works since I've seen quite some reports from it ;)

								Honza
-- 
Jan Kara <jack at suse.cz>
SUSE Labs, CR