On Tue, Jun 16, 2020 at 08:17:28AM -0400, Bob Peterson wrote: > ----- Original Message ----- > > > I'd assume Andreas is looking at converting a filesystem to use iomap, > > > since this problem only occurs for filesystems which have returned an > > > invalid extent. > > > > Well, I can assume it's gfs2, but you know what happens when you > > assume something.... > > Yes, it's gfs2, which already has iomap. I found the bug while just browsing > the code: gfs2 takes a lock in the begin code. If there's an error, > however unlikely, the end code is never called, so we would never unlock. > It doesn't matter to me whether the error is -EIO because it's very unlikely > in the first place. I haven't looked back to see where the problem was > introduced, but I suspect it should be ported back to stable releases. It shouldn't just be "unlikely", it should be impossible. This is the iomap code checking whether you've returned an extent which doesn't cover the range asked for. I don't think it needs to be backported, and I'm pretty neutral on whether it needs to be applied.
