On Wed 21-01-15 17:07:57, Jan Kara wrote:
> Hi Dave,
>
> On Wed 21-01-15 11:39:37, Dave Chinner wrote:
> > These three patches detect and fix the issues you reported with log
> > recovery finding buffers with a format type of zero. The type of
> > zero (XFS_BLFT_UNKNOWN_BUF) is only valid for buffers that have
> > been cancelled (i.e. invalidated or marked stale as they have been
> > freed), so the series adds asserts to ensure these conditions are
> > met during transaction commit. Hence we shouldn't ever get new code
> > that fails to set the buffer type getting through testing.
> >
> > The last two patches fix the cases that running xfstests uncovered
> > where we don't set the buffer type appropriately. There may be more,
> > but doing this much made my head hurt and xfstests is clean, so it's
> > as much as I'm going to do right now. Can you test it and see if it
> > runs clean (with CONFIG_XFS_WARN=y or CONFIG_XFS_DEBUG=y) on your
> > test setup?
> Thanks for a quick response with patches. My test round has finished and
> the new assertion didn't trigger so things look fine. I've also provided a
> test kernel to the guy seeing these issues in the wild but there it took
> days / weeks to trigger so I wouldn't wait for it...
BTW, this should be stable material I guess (just that I didn't see CC to
stable in the patches).
Honza
--
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR
|