Re: [PATCH v3 2/8] xfs: flush dirty pagecache over hole in zoned mode zero range

[Brian Foster <bfoster@redhat.com>]

On Mon, Mar 09, 2026 at 10:22:19AM -0700, Darrick J. Wong wrote:
> On Mon, Mar 09, 2026 at 09:45:00AM -0400, Brian Foster wrote:
> > For zoned filesystems a window exists between the first write to a
> > sparse range (i.e. data fork hole) and writeback completion where we
> > might spuriously observe holes in both the COW and data forks. This
> > occurs because a buffered write populates the COW fork with
> > delalloc, writeback submission removes the COW fork delalloc blocks
> > and unlocks the inode, and then writeback completion remaps the
> > physically allocated blocks into the data fork. If a zero range
> > operation does a lookup during this window where both forks show a
> > hole, it incorrectly reports a hole mapping for a range that
> > contains data.
> > 
> > This currently works because iomap checks for dirty pagecache over
> > holes and unwritten mappings. If found, it flushes and retries the
> > lookup. We plan to remove the hole flush logic from iomap, however,
> > so lift the flush into xfs_zoned_buffered_write_iomap_begin() to
> > preserve behavior and document the purpose for it. Zoned XFS
> > filesystems don't support unwritten extents, so if zoned mode can
> > come up with a way to close this transient hole window in the
> > future, this flush can likely be removed.
> 
> Why does the mapping disappear out of both data and cow forks between
> writeback setup and completion?  IIRC it is because the writeback ioend
> effectively owns the unwritten mapping.  We want another writer thread
> to see the hole and reserve its own out-of-place write because the write
> mapping that writeback's working on is immutable once the disk actually
> writes it.  Right?
> 

Not sure.. seems plausible, but probably a question for Christoph.

> I wonder if we could stash a delalloc mapping in the cow fork with zero
> indlen during writeback to signal "get a real zoned space reservation"?
> 

I was thinking something like transferring the delalloc to the data fork
when the latter has a hole just to signify that there is data in the
pipeline, but not something to solve for this series anyways..

> > Signed-off-by: Brian Foster <bfoster@redhat.com>
> > ---
> >  fs/xfs/xfs_iomap.c | 19 +++++++++++++++++++
> >  1 file changed, 19 insertions(+)
> > 
> > diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> > index 8c3469d2c73e..0e323e4e304b 100644
> > --- a/fs/xfs/xfs_iomap.c
> > +++ b/fs/xfs/xfs_iomap.c
> > @@ -1590,6 +1590,7 @@ xfs_zoned_buffered_write_iomap_begin(
> >  {
> >  	struct iomap_iter	*iter =
> >  		container_of(iomap, struct iomap_iter, iomap);
> > +	struct address_space	*mapping = inode->i_mapping;
> >  	struct xfs_zone_alloc_ctx *ac = iter->private;
> >  	struct xfs_inode	*ip = XFS_I(inode);
> >  	struct xfs_mount	*mp = ip->i_mount;
> > @@ -1614,6 +1615,7 @@ xfs_zoned_buffered_write_iomap_begin(
> >  	if (error)
> >  		return error;
> >  
> > +restart:
> >  	error = xfs_ilock_for_iomap(ip, flags, &lockmode);
> >  	if (error)
> >  		return error;
> > @@ -1686,8 +1688,25 @@ xfs_zoned_buffered_write_iomap_begin(
> >  	 * When zeroing, don't allocate blocks for holes as they are already
> >  	 * zeroes, but we need to ensure that no extents exist in both the data
> >  	 * and COW fork to ensure this really is a hole.
> > +	 *
> > +	 * A window exists where we might observe a hole in both forks with
> > +	 * valid data in cache. Writeback removes the COW fork blocks on
> > +	 * submission but doesn't remap into the data fork until completion. If
> > +	 * the data fork was previously a hole, we'll fail to zero. Until we
> > +	 * find a way to avoid this transient state, check for dirty pagecache
> > +	 * and flush to wait on blocks to land in the data fork.
> >  	 */
> >  	if ((flags & IOMAP_ZERO) && srcmap->type == IOMAP_HOLE) {
> > +		if (filemap_range_needs_writeback(mapping, offset,
> > +						  offset + count - 1)) {
> > +			xfs_iunlock(ip, lockmode);
> > +			error = filemap_write_and_wait_range(mapping, offset,
> > +							    offset + count - 1);
> 
> Two tab indents, please.
> 

Sure.

Brian

> --D
> 
> > +			if (error)
> > +				return error;
> > +			goto restart;
> > +		}
> > +
> >  		xfs_hole_to_iomap(ip, iomap, offset_fsb, end_fsb);
> >  		goto out_unlock;
> >  	}
> > -- 
> > 2.52.0
> > 
> > 
>