On Wed, Aug 21, 2024 at 08:44:31AM -0400, Brian Foster wrote:

quoted

+	error = xfs_reflink_unshare(XFS_I(inode), offset, len);
+	if (error)
+		return error;
+

Doesn't unshare imply alloc?

Yes, ooks like that got lost and no test noticed it

quoted

-	if (xfs_file_sync_writes(file))
+	if (!error && xfs_file_sync_writes(file))
 		error = xfs_log_force_inode(ip);

I'd think if you hit -ENOSPC or something after doing a partial alloc to
a sync inode, you'd still want to flush the changes that were made..?

Persistence behavior on error is always undefined.  And that's also
what the current code does, as it jumps past the log force from all
error exits.