Re: [PATCH 3/3] xfs: cancel COW in xfs_cancel_ioend

[Christoph Hellwig <hch@xxxxxx>]

On Sat, Jan 09, 2016 at 01:17:13PM -0800, Darrick J. Wong wrote:
> > 
> > I've not seen the inode eviction asserts anymore, but I now hit a
> > corruption warnings in generic/168 reliably.  I did hit before as
> > well, but not very reliably.
> 
> I'll see if I can repro the 168 error; it's been running in a loop all
> night and hasn't bombed yet.

Note that this is over nfs to a local server, not running on xfs directly,
which is doing fine.

> In the meantime, I added some more tests and fixed a CoW corruption when an
> xfs_io_overwrite extent has cow reservations in the middle of the extent.
> 
> I also restarted testing on arm64, ppc64{,el}, and i686; it seems
> stable enough right now to pass all ~130 reflink xfstests here.

I see pretty reliable failures in xfs/128 xfs/132 xfs/139, apparenly
due to content mismatches.

Re the verifier failure:

sees like we're hitting the

        if (level >= pag->pagf_refcount_level)
                return false;

case.  Together with the other garbage in it seems like we're
seeing a btree block that's not properly initialized in some way,
maybe after a split.