On Fri, Aug 20, 2021 at 08:18:44AM -0700, Dan Williams wrote:

I notice that the @iomap argument to ->iomap_end() is reliably coming
from @iter. So you could do the following in your iomap_end()
callback:

        struct iomap_iter *iter = container_of(iomap, typeof(*iter), iomap);
        struct xfs_inode *ip = XFS_I(inode);
        ssize_t written = iter->processed;
        bool cow = xfs_is_cow_inode(ip);

        if (cow) {
                if (written <= 0)
                        xfs_reflink_cancel_cow_range(ip, pos, length, true)
        }

I think this might be ok for now (with a big comment).  Willy's original
iomap iter series replaced the iomap_begin and iomap_end with a single
next callback that takes the iomap_iter, which would solve this issue.
My plan is to look into a series that implements this and sees if this
is indeed a net win.