watch_queue: Fix lack of barrier/sync/lock between post and read

commit 2ed147f upstream.

There's nothing to synchronise post_one_notification() versus
pipe_read().  Whilst posting is done under pipe->rd_wait.lock, the
reader only takes pipe->mutex which cannot bar notification posting as
that may need to be made from contexts that cannot sleep.

Fix this by setting pipe->head with a barrier in post_one_notification()
and reading pipe->head with a barrier in pipe_read().

If that's not sufficient, the rd_wait.lock will need to be taken,
possibly in a ->confirm() op so that it only applies to notifications.
The lock would, however, have to be dropped before copy_page_to_iter()
is invoked.

Fixes: c73be61 ("pipe: Add general notification queue support")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

fs/pipe.c

@@ -252,7 +252,8 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 	 */
 	was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 	for (;;) {
-		unsigned int head = pipe->head;
+		/* Read ->head with a barrier vs post_one_notification() */
+		unsigned int head = smp_load_acquire(&pipe->head);
 		unsigned int tail = pipe->tail;
 		unsigned int mask = pipe->ring_size - 1;
 

kernel/watch_queue.c

@@ -113,7 +113,7 @@ static bool post_one_notification(struct watch_queue *wqueue,
 	buf->offset = offset;
 	buf->len = len;
 	buf->flags = PIPE_BUF_FLAG_WHOLE;
-	pipe->head = head + 1;
+	smp_store_release(&pipe->head, head + 1); /* vs pipe_read() */
 
 	if (!test_and_clear_bit(note, wqueue->notes_bitmap)) {
 		spin_unlock_irq(&pipe->rd_wait.lock);