Re: [PATCH 1/2] [PATCH 1/2] direct-io: implement generic deferred AIO completions

[Jan Kara <jack@suse.cz>]

On Sat 08-12-12 07:02:03, Christoph Hellwig wrote:
> On Thu, Dec 06, 2012 at 08:05:37PM +0100, Jan Kara wrote:
> >   I like this, but this patch already breaks ext4, doesn't it?
> 
> >   ^^^ Here ext4 offloads IO completion to a worker thread. So you now
> > complete AIO / DIO before ext4_end_io() runs which is a bug (ext4_end_io()
> > is responsible for example for calling end_page_writeback()). I'll modify
> > these patches to work for ext4 tomorrow I hope...
> 
> You're right, patch 2 actively deadlocks ext4 under xfstests, but the
> first one already breaks semantics.  So any rework of the ext4 unwritten
> extent handling really should go before these two patches.
  Just to let you know, I didn't forget about this. I even have patches
to change PageWriteback handling in ext4 but I realized it can cause
deadlocks.

When converting unwritten extents to written ones, we need to start a
transaction, that may block because of lack of journal space waiting for
currently running transaction to finish. And that running transaction never
finishes because someone (e.g. ext4_write_begin()) has a handle to it waiting
for PageWriteback bit (e.g. in grab_cache_page_write_begin()).

I'm pondering how we could fix this in ext4 because I'd really like to get
rid of that PageWriteback handling oddity in ext4. Just it's not simple...

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