Re: yet another approach to fix the loop lock order inversions v6

[Jan Kara <jack@suse.cz>]

On Tue 05-04-22 08:28:38, Christoph Hellwig wrote:
> On Mon, Apr 04, 2022 at 06:39:31PM +0900, Tetsuo Handa wrote:
> > Two bugs which Jan has found in /bin/mount might not be yet fixed in
> > versions developers/users are using. Thus, let's wait for a while
> > before committing to linux.git.
> 
> Jan, which loop bugs might be relevant here?

So there was a bug in libmount code trying to reuse already setup loop
devices and changes in timing & removing the lo_mutex from lo_open() +
suitable racing with systemd-udev probing devices caused these bugs to
manifest. The visible effect was that a loop like:

while :; do mount -o loop,ro isofs.iso isofs/; umount isofs/; done

often failed with:

mount: /mnt: can't read superblock on /dev/loop0.
umount: /mnt: not mounted.

Bugs are fixed now by commits 3e1fc3bbe ("mount: Fix race in loop device
reuse code"), fb4b6b115 ("loopdev: Properly translate errors from
ul_path_read_*()"), 562990b552 ("loopdev: Do not treat errors when
detecting overlap as fatal") in util-linux git.

The only real-world impact I've heard about was LTP failing due to these
problems and those guys should be capable enough to update their util-linux
when we tell them. So I think we should be OK with pushing the kernel fixes
upstream but it may generate some noise from automated testers before
util-linux is updated on all of them...

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR