Mounting on floating mounts is possible

Hi,

this[1] patch changed check_mnt() so that mounting on a floating mount (i.e. one that was unmounted using MNT_DETACH and was still in use) is possible, since we no longer check if the mountpoint is actually reachable. The problem is that we may lose any reference to the floating mount, but the mount on it will keep it alive, thus it will never go away. The following patch removes the reference from the mount to its namespace when it is unmounted lazily, so that check_mnt protects from such mounts.

Please CC me as I'm not subscribed to the list.

Bjoern

[1] http://lwn.net/Articles/91946/

diff -uNr --minimal a/fs/namespace.c b/fs/namespace.c

Quoted 17 lines · Björn Steinbrink

Hi,

this[1] patch changed check_mnt() so that mounting on a floating mount(i.e. one that was unmounted using MNT_DETACH and was still in use) is possible, since we no longer check if the mountpoint is actually reachable. The problem is that we may lose any reference to the floating mount, but the mount on it will keep it alive, thus it will never go away. The following patch removes the reference from the mount to its namespace when it is unmounted lazily, so that check_mnt protects from such mounts.

Please CC me as I'm not subscribed to the list.

Bjoern

[1] http://lwn.net/Articles/91946/

fs/namespace.c 2004-11-06 04:38:37.299013810 +0100

Quoted 12 lines

diff -uNr --minimal a/fs/namespace.c b/fs/namespace.c

fs/namespace.c 2004-11-06 04:38:37.299013810 +0100 renamed from fs/namespace.c 2004-10-31 00:41:02.000000000 +0200 +1 -1

@@ -358,6 +358,7 @@
                } else {
                        struct nameidata old_nd;
                        detach_mnt(mnt, &old_nd);
+                       mnt->mnt_namespace = NULL;
                        spin_unlock(&vfsmount_lock);
                        path_release(&old_nd);
                }
-

I don't think this patch clears mnt_namespace for the root of the umounted tree. How about this?

Quoted 19 lines · On Mon, 08 Nov 2004 11:20:27 -0500 Mike Waychison <[email protected]>…

Quoted 17 lines · Björn Steinbrink

Hi,

this[1] patch changed check_mnt() so that mounting on a floating mount(i.e. one that was unmounted using MNT_DETACH and was still in use) is possible, since we no longer check if the mountpoint is actually reachable. The problem is that we may lose any reference to the floating mount, but the mount on it will keep it alive, thus it will never go away. The following patch removes the reference from the mount to its namespace when it is unmounted lazily, so that check_mnt protects from such mounts.

Please CC me as I'm not subscribed to the list.

Bjoern

[1] http://lwn.net/Articles/91946/

Quoted 16 lines

fs/namespace.c 2004-11-06 04:38:37.299013810 +0100

Quoted 12 lines

diff -uNr --minimal a/fs/namespace.c b/fs/namespace.c

fs/namespace.c 2004-11-06 04:38:37.299013810 +0100 renamed from fs/namespace.c 2004-10-31 00:41:02.000000000 +0200 +1 -1

@@ -358,6 +358,7 @@
               } else {
                       struct nameidata old_nd;
                       detach_mnt(mnt, &old_nd);
+                       mnt->mnt_namespace = NULL;
                       spin_unlock(&vfsmount_lock);
                       path_release(&old_nd);
               }
-

I don't think this patch clears mnt_namespace for the root of the umounted tree. How about this?

The root mount of the umounted tree is not detached yet, i.e. it has mnt->mnt_parent != mnt and is detached in the else-branch, so that is handled fine (at least my tests said so ;).
Only detached mounts and the namespace's root mount (that's the rootfs mount IIRC) have mnt->mnt_parent==mnt. And we have two cases here:
a) already detached mount, this cannot happen. The only possibility to get a detached mount into umount is the same we currently have with mount, some process kept a reference and used a relative path. But with the patch we would already bail out in check_mnt earlier in this case.
b) root mount, is not detached so all mounts are still reachable and can be unmounted as usual.