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

[Darrick J. Wong]

On Mon, Jan 04, 2016 at 05:43:10PM -0800, Darrick J. Wong wrote:
> On Sun, Jan 03, 2016 at 01:07:53PM +0100, Christoph Hellwig wrote:
> > Otherwise we leak COW allocations done earlier in writepage.  This
> > can be reproduced fairly easily when we hit the non-blocking writeback
> > EAGAIN case.
> > 
> > Signed-off-by: Christoph Hellwig <hch at lst.de>
> > ---
> >  fs/xfs/xfs_aops.c | 7 +++++++
> >  1 file changed, 7 insertions(+)
> > 
> > diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
> > index 185415a..9c69dc3 100644
> > --- a/fs/xfs/xfs_aops.c
> > +++ b/fs/xfs/xfs_aops.c
> > @@ -588,6 +588,7 @@ xfs_cancel_ioend(
> >  {
> >  	xfs_ioend_t		*next;
> >  	struct buffer_head	*bh, *next_bh;
> > +	int			error;
> >  
> >  	do {
> >  		next = ioend->io_list;
> > @@ -605,6 +606,12 @@ xfs_cancel_ioend(
> >  			unlock_buffer(bh);
> >  		} while ((bh = next_bh) != NULL);
> >  
> > +		if (ioend->io_flags & XFS_IOEND_COW) {
> > +			error = xfs_reflink_end_cow_failed(
> > +					XFS_I(ioend->io_inode),
> > +					ioend->io_offset, ioend->io_size);
> > +			WARN_ON_ONCE(error);
> > +		}

Actually, no, this isn't correct.  Even if we cancel the ioend, we must retain
the CoW reservation because the pages remain dirty and writepage will try
again.  If we delete the reservation, that second writepage will treat the
dirty page as a regular overwrite because there's no reservation, which is
wrong.

We need to keep something in the CoW fork; either we can leave the allocated
blocks or we could theoretically convert it back to a delalloc reservation.
For now I'll leave the mapping untouched since I've subsequently taught xfs to
clear out the CoW mappings when we truncate/punch/etc.  The reservation won't
hang around for long unless IO errors start piling up.

This causes file corruption in xfs/140 when blocksize < pagesize.

--D

> 
> Hmm.  This might be the cause of the occasional complaints I've been seeing
> where allocated blocks remain in the COW fork when the inode is being cleared
> out.  That said, the xfs_reflink_end_cow_failed() is apparently missing a
> xfs_bunmapi_cow() to actually clean out the COW fork.
> 
> Good catch, in any case.  Thank you for the testing and patches! :)
> 
> --D
> 
> >  		mempool_free(ioend, xfs_ioend_pool);
> >  	} while ((ioend = next) != NULL);
> >  }
> > -- 
> > 1.9.1
> > 
> > _______________________________________________
> > xfs mailing list
> > xfs at oss.sgi.com
> > http://oss.sgi.com/mailman/listinfo/xfs
> 
> _______________________________________________
> xfs mailing list
> xfs at oss.sgi.com
> http://oss.sgi.com/mailman/listinfo/xfs