Re: [PATCH] xfs: fix XFS_ERRTAG_FORCE_ZERO_RANGE for zoned file system

[Brian Foster <bfoster@redhat.com>]

On Wed, Dec 10, 2025 at 04:40:16PM +0100, Christoph Hellwig wrote:
> On Wed, Dec 10, 2025 at 10:36:55AM -0500, Brian Foster wrote:
> > Is there a reason in particular for testing this with the zone mode?
> > It's just a DEBUG thing for the zeroing mechanism. Why not just filter
> > out the is_zoned_inode() case at the injection site?
> 
> Because I also want to be able to test the zeroing code for zoned
> file systems, especially given zeroing is a bit of painful area
> for out of place write file systems like zoned XFS.
> 
> > I suppose you could argue there is a point if we have separate zoned
> > mode iomap callbacks and whatnot, but I agree the factoring here is a
> > little unfortunate. I wonder if it would be nicer if we could set a flag
> > or something on an ac and toggle the zone mode off that, but on a quick
> > look I don't see a flag field in the zone ctx.
> 
> I don't really follow what you mean here.
> 

I was just rambling about if/how we might be able to use the ac..

> > Hmm.. I wonder if we could still do something more clever where the zone
> > mode has its own injection site to bump the res, and then the lower
> > level logic just checks whether the reservation is sufficient for a full
> > zero..? I'm not totally sure if that's ultimately cleaner, but maybe
> > worth a thought..
> 
> We could have a different site for that injection, but we'd still need
> to move the current one or at least make it conditional so that it
> can't trigger for zoned mode.  I doubt that's less ugly than this
> version.
> 

Well yeah, it would look something like this at the current site:

	if (!is_inode_zoned() && XFS_TEST_ERROR(...) ||
	    ac->reserved_blocks == magic_default_res + len)
		xfs_zero_range(...);
	else
		xfs_free_file_space(...);

... and the higher level zoned code would clone the XFS_TEST_ERROR() to
create the block reservation condition to trigger it.

Alternatively perhaps you could make that check look something like:

	if (XFS_TEST_ERROR() && (!ac || ac->res > len))
		...
	else
		...

... and let the res side always bump the res in DEBUG mode, with a
fallback on -ENOSPC or something.

Actually the latter sounds potentially more clean to me, but I don't
object to this if not.

Brian