On Mon, Jun 10, 2013 at 07:17:22PM -0400, Brian Foster wrote:
> On 06/10/2013 05:31 PM, Dave Chinner wrote:
> > On Mon, Jun 10, 2013 at 09:17:31AM -0400, Brian Foster wrote:
> >> Hi guys,
> >> I wanted to get this onto the list... I suspect this could be
> >> similar/related to the issue reported here:
> >> http://oss.sgi.com/archives/xfs/2013-06/msg00066.html
> > Unlikely - generic/263 tests mmap IO vs direct IO, and Sage's
> > problem has neither...
> Oh, Ok. I didn't look at that one closely enough then.
> >> While running xfstests, the only apparent regression I hit from 3.9.0
> >> was generic/263. This test fails due to the following command (and
> >> resulting output):
> > Not a regression - 263 has been failing ever since it was introduced
> > in 2011 by:
> > commit 0d69e10ed15b01397e8c6fd7833fa3c2970ec024
> > It is testing mmap() writes vs direct IO, something that is known to
> > be fundamentally broken (i.e. racy) as mmap() page fault path does
> > not hold the XFS_IOLOCK or i_mutex in any way. The direct IO path
> > tries to wark around this by flushing and invalidating cached pages
> > before IO submission, but the lack of locking in the page fault path
> > means we can't avoid the race entirely.
> Thanks for the explanation.
> >> P.S., I also came across the following thread which, if related,
> >> suggests this might be known/understood to a degree:
> >> http://oss.sgi.com/archives/xfs/2012-04/msg00703.html
> > Yup, that's potentially one aspect of it. However, have you run the
> > test code on ext3/4? it works just fine - it's only XFS that has
> > problems with this case, so it's not clear that this is a DIO
> > problem. It was never able to work out where ext3/ext4 were zeroing
> > the part of the page beyond EOF, and I couldn't ever make the DIO
> > code reliably do the right thing. It's one of the reasons that lead
> > to this discussion as LSFMM:
> > http://lwn.net/Articles/548351/
> Interesting, thanks again. I did happen to run the script and the fsx
> test on the ext4 rootfs of my VM and observed expected behavior.
> Note that I mentioned this was harder to reproduce with fixed alloc
> sizes less than 128k or so. I don't believe ext4 does any kind of
> speculative preallocation in the manner that XFS does. Perhaps that is a
Oh, it most likely is, but XFS has done speculative prealloc since,
well, forever, so this isn't a regression as such. FWIW, the old
default for speculative prealloc was XFS_WRITEIO_LOG_LARGE (16
filesystem blocks), so this test would have failed before any of the
dynamic speculative alloc changes were made....
Indeed, if you mount with -o allocsize=4k, you'll find the test case
no longer fails - it requires allocsize=32k (or larger) to fail
here. That's not surprising, given that the test is writing across a
16k-beyond-eof boundary when it triggers the problem, and so needs a
prealloc size of >16k to trigger...