cachefiles: fix the race between cachefiles_bury_object() and rmdir(2)

commit 169b803397499be85bdd1e3d07d6f5e3d4bd669e upstream.

the victim might've been rmdir'ed just before the lock_rename();
unlike the normal callers, we do not look the source up after the
parents are locked - we know it beforehand and just recheck that it's
still the child of what used to be its parent.  Unfortunately,
the check is too weak - we don't spot a dead directory since its
->d_parent is unchanged, dentry is positive, etc.  So we sail all
the way to ->rename(), with hosting filesystems _not_ expecting
to be asked renaming an rmdir'ed subdirectory.

The fix is easy, fortunately - the lock on parent is sufficient for
making IS_DEADDIR() on child safe.

Cc: stable@vger.kernel.org
Fixes: 9ae326a690 (CacheFiles: A cache that backs onto a mounted filesystem)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

fs/cachefiles/namei.c

@@ -323,7 +323,7 @@ try_again:
 	trap = lock_rename(cache->graveyard, dir);
 
 	/* do some checks before getting the grave dentry */
-	if (rep->d_parent != dir) {
+	if (rep->d_parent != dir || IS_DEADDIR(d_inode(rep))) {
 		/* the entry was probably culled when we dropped the parent dir
 		 * lock */
 		unlock_rename(cache->graveyard, dir);