On Fri, Jul 12, 2013 at 10:44:21AM +1000, Dave Chinner wrote:
> On Thu, Jul 11, 2013 at 12:02:18AM +0200, Jan Kara wrote:
> > From: Christoph Hellwig <hch@xxxxxxxxxxxxx>
> > Add support to the core direct-io code to defer AIO completions to user
> > context using a workqueue. This replaces opencoded and less efficient
> > code in XFS and ext4 and will be needed to properly support O_(D)SYNC
> > for AIO.
> I don't see how this is any more efficient than the deferral
> that XFS already does for direct IO completion - it just changes
> what queues IO for completions. And on that topic:
It avoided a memory allocation for each dÑrect I/O write.
> FWIW, in XFS we queue unwritten extent conversion completions on a
> different workqueue to EOF size update completions because the
> latter are small, fast and rarely require IO or get blocked. The
> data IO completion workqueue for EOF updates has the same
> concurrency and depth as the unwritten extent work queue (i.e. 256
> workers per cpu per superblock). So pushing all of this DIO and EOF
> completion work into a single threaded global workqueue that can
> block in every IO completion doesn't seem like a very smart idea to
Currently the direct I/O code doesn't support async extending writes.