On Fri, Feb 21, 2020 at 5:33 PM David Howells [off-list ref] wrote:

Jann Horn [off-list ref] wrote:

quoted

(And as in the other case, the s->s_count increment will probably have
to be moved above the add_watch_to_object(), unless you hold the
sb_lock around it?)

It shouldn't matter as I'm holding s->s_umount across the add and increment.
That prevents the watch from being removed: watch_sb() would have to get the
lock first to do that.  It also deactivate_locked_super() from removing all
the watchers.

Can't the same thing I already pointed out on "[PATCH 13/19] vfs: Add
a mount-notification facility [ver #16]" also happen here?

If another thread concurrently runs close(watch_fd) before the
spin_lock(&sb_lock), pipe_release -> put_pipe_info -> free_pipe_info
-> watch_queue_clear will run, correct? And then watch_queue_clear()
will find the watch that we've just created and call its
->release_watch() handler, which causes put_super(), potentially
dropping the refcount to zero? And then stuff will blow up.

I can move it before, but I probably have to drop s_umount before I can call
put_super().