On Wed, Nov 28, 2012 at 11:07:36AM -0500, Christoph Hellwig wrote:
> On Wed, Nov 28, 2012 at 01:12:48PM +1100, Dave Chinner wrote:
> > From: Dave Chinner <dchinner@xxxxxxxxxx>
> > lockdep reports splice vs direct-io write lock inversions due to
> > generic_file_splice_write() taking the inode->i_mutex inside
> > XFS_IOLOCK_EXCL context. These lock contexts are inverted, hence can
> > deadlock. Remove the XFS_IOLOCK_EXCL locking context from the outer
> > function and drive it inwards to the actor function that only locks
> > the inode when the lock is really needed,
> Otherwise the patch looks fine, but I'd love to understand why the
> generic code thes te I_MUTEX_CHILD annotation and we can get away
> without it.
> Also can you add a testcase for this to xfstests?
I've had trouble reproducing it reliably. The only cases I've seen
it occur are some custom FIO workloads. I'll try again to see if I
can come up with something that reliably hits it...