On Wed, Feb 19, 2020 at 11:40 PM Jann Horn [off-list ref] wrote:

On Tue, Feb 18, 2020 at 6:07 PM David Howells [off-list ref] wrote:

[...]

quoted

+                       watch = NULL;
+               }
+               up_write(&m->mnt.mnt_sb->s_umount);
+       } else {
+               ret = -EBADSLT;
+               if (m->mnt_watchers) {
+                       down_write(&m->mnt.mnt_sb->s_umount);
+                       ret = remove_watch_from_object(m->mnt_watchers, wqueue,
+                                                      (unsigned long)path.dentry,
+                                                      false);

What exactly is the implication of only using the dentry as key here
(and not the mount)? Does this mean that if you watch install watches
on two bind mounts of the same underlying filesystem, the notification
mechanism gets confused?

Ah, nevermind, I understand this one now... this operation only
removes watches on this mount with that dentry, and so together, that
means it effectively removes watches by the full path.