On Thu, Apr 10, 2014 at 03:00:47PM +1000, Dave Chinner wrote:
> Hi folks,
> This is version 2 of the DIO vs delalloc patchset I posted here:
> The changes to this version are:
> - the bug fix to patch 2 that Brain noticed,
> - I dropped the delalloc extent splittting patch because
> with the fix to patch 2 I can't trigger that bug anymore,
> and that patch was causing transaction overruns in
> xfs/297. Hence without an existing reproducer, I won't try
> to fix that problem.
> - the last patch is new, and is a bug in the collapse range
> code where it fails to shift the last N extents correctly
> if there are N delalloc extents before the shifted range.
> With these 6 patches, all of the xfstests fsx/fsstress tests pass
> on 1k, 2k and 4k block size filesystems, with and without CRCs
> enabled, on 1, 2 and 16p test VMs.
> I'm much happier with these patches now - I don't think that there
> are more problems lurking, but only time will tell. I'd like to get
> these fixes to Linus for 3.15 (probably for -rc2), so eyeballs and
> testing would be appreciated.
The series looks pretty good to me now with the latest fix. I was
previously tripping all over the delalloc asserts. With this set (and
also running with finobt enabled), my tests ran clean on a 4k fs.
I've seen generic/270 still hit an assert once or twice on a 1k block fs
(as noted on irc), but that's proven rather difficult to reproduce. I'll
continue beating on it a bit, but otherwise:
Tested-by: Brian Foster <bfoster@xxxxxxxxxx>
> xfs mailing list