Re: [RFC] Badness in __mutex_unlock_slowpath with XFS stress tests

[Nathan Scott <nathans@xxxxxxx>]

On Thu, Mar 09, 2006 at 12:03:06PM +0000, Christoph Hellwig wrote:
> On Thu, Mar 09, 2006 at 01:24:38PM +0530, Suzuki wrote:
> > 
> > Missed out linux-aio & linux-fs-devel lists. Forwarding.
> > 
> > Comments ?
> 
> I've seen this too.  The problem is that __generic_file_aio_read can return
> with or without the i_mutex locked in the direct I/O case for filesystems
> that set DIO_OWN_LOCKING.

Not for reads AFAICT - __generic_file_aio_read + own-locking
should always have released i_mutex at the end of the direct
read - are you thinking of writes or have I missed something?

> It's a nasty one and I haven't found a better solution
> than copying lots of code from filemap.c into xfs.

Er, eek?  Hopefully thats not needed - from my reading of the
code, all the i_mutex locking for direct reads lives inside
direct-io.c, not filemap.c -- is the solution from my other
mail not workable?  (isn't it only writes that has the wierd
buffered I/O fallback + i_sem/i_mutex locking interaction?).

thanks.

-- 
Nathan