Re: [PATCH v4] fs/namespace: defer RCU sync for MNT_DETACH umount

[Al Viro <viro@zeniv.linux.org.uk>]

On Thu, Apr 17, 2025 at 05:12:26PM +0200, Christian Brauner wrote:

> (2) If a userspace task is dealing with e.g., a broken NFS server and
>     does a umount(MNT_DETACH) and that NFS server blocks indefinitely
>     then right now it will be the task's problem that called the umount.
>     It will simply hang and pay the price.
> 
>     With your patch however, that cleanup_mnt() and the
>     deactivate_super() call it entails will be done from
>     delayed_mntput_work...
> 
>     So if there's some userspace process with a broken NFS server and it
>     does umount(MNT_DETACH) it will end up hanging every other
>     umount(MNT_DETACH) on the system because the dealyed_mntput_work
>     workqueue (to my understanding) cannot make progress.
> 
>     So in essence this patch to me seems like handing a DOS vector for
>     MNT_DETACH to userspace.

(3) Somebody does umount -l and a few minutes later proceeds to reboot.
All filesystems involved are pinned only by mounts, but the very first
victim happens to be an NFS mount from a slow server.  No indication
of the problem, just a bunch of local filesystems that got a dirty shutdown...