On Tue, Jun 26, 2012 at 12:07:17AM -0600, Miao Xie wrote:

The 209th case of xfstests failed because of the race between aio and dio. The
detail reason is following:
	Task1		Task2			Btrfs-worker
			invalidate pages
	read pages
			do direct io
			invalidate pages fail*
						finish ordered io
	read data from
	pages

This just papers over the problem and makes DIO touch page cache which it
shouldn't be doing if it's working properly, so NAK.  We need to figure out why
exactly my patch didn't work, since it should be working.  The write should be
doing

lock_extent
setup ordered extent
unlock_extent

and the read should be doing

lock_extent
check for ordered extent
	if there is one unlock and wait and then loop
do read
unlock_extent

there should be no room for races in here.  The patch I sent earlier should have
caught if we had done a read between the invalidate and the locking and should
be invalidating the range again and then checking.  If this isn't working then
something else is going sideways and we really need to figure out what it is
rather than just working around the issue, as it will likely bite us in a
different way later.  Thanks,

Josef