Re: [PATCH] ext4: fix race between truncate and __ext4_journalled_writepage()

[Jan Kara <jack@suse.cz>]

On Mon 15-06-15 09:06:11, Ted Tso wrote:
> On Mon, Jun 15, 2015 at 02:33:52PM +0200, Jan Kara wrote:
> >   Yeah, that's nasty. Thanks for debugging this! However I think your fix
> > reintroduces the original deadlock issues. do_journal_get_write_access()
> > can end up blocking waiting for jbd2 thread to finish a commit while jbd2
> > thread may be blocked waiting for the page to be unlocked.
> > 
> > After some thought I don't think the deadlock is real since
> > do_journal_get_write_access() will currently only block if a buffer is
> > under writeout to the journal and at that point we don't wait for page
> > locks anymore. Also ext4_write_begin() does the same in data=journal mode
> > and we haven't observed deadlocks so far. But still things look really
> > fragile here.
> 
> The reason why there are no deadlocks is the writeback in the commit
> thread happens when the inode gets written back --- but that only
> happens for data=ordered inodes, not data=journalled mode.  I was a
> little worried about what might happen when after the 'j' chattr
> attribute gets set on an inode, and the inode was still on the ordered
> flush list.
> 
> Hmm... I think we could also maybe fix this by having
> ext4_change_inode_journal_flag() force a journal commit before setting
> the JOURNAL_DATA flag.  If we did that, we could just avoid dropping
> the page_lock in __ext4_journalled_writepage() altogether.
> 
> What do you think?
  I think that fully switching lock ordering for data=journal mode back to
page lock -> transaction start (which is what you effectively do when you
never drop page lock in ->writepage) is rather error prone. We'd have to be
careful to avoid lock inversion also for places like ->write_begin,
->releasepage, ->invalidatepage etc. For example ext4_write_begin() will
currently call lock_page() with transaction started which could deadlock
against journalled writepage you suggest. So effectively we'd have to
completely separate aops for data=journal mode. Doable but I'm not sure
it's worth it.

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