poll_wait: add mb() to fix theoretical race between waitqueue_active() and .poll()

As the comment above waitqueue_active() explains, it can only be used
if both waker and waiter have mb()'s that pair with each other. However
__pollwait() is broken in this respect.

This is not pipe-specific, but let's look at pipe_poll() for example:

	poll_wait(...); // -> __pollwait() -> add_wait_queue()

	LOAD(pipe->head);
	LOAD(pipe->head);

In theory these LOAD()'s can leak into the critical section inside
add_wait_queue() and can happen before list_add(entry, wq_head), in this
case pipe_poll() can race with wakeup_pipe_readers/writers which do

	smp_mb();
	if (waitqueue_active(wq_head))
		wake_up_interruptible(wq_head);

There are more __pollwait()-like functions (grep init_poll_funcptr), and
it seems that at least ep_ptable_queue_proc() has the same problem, so the
patch adds smp_mb() into poll_wait().

Link: https://lore.kernel.org/all/20250102163320.GA17691@redhat.com/
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Link: https://lore.kernel.org/r/20250107162717.GA18922@redhat.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
This commit is contained in:
Oleg Nesterov 2025-01-07 17:27:17 +01:00 committed by Christian Brauner
parent 40384c840e
commit cacd9ae4bf
No known key found for this signature in database
GPG Key ID: 91C61BC06578DCA2

View File

@ -41,8 +41,16 @@ typedef struct poll_table_struct {
static inline void poll_wait(struct file * filp, wait_queue_head_t * wait_address, poll_table *p)
{
if (p && p->_qproc && wait_address)
if (p && p->_qproc && wait_address) {
p->_qproc(filp, wait_address, p);
/*
* This memory barrier is paired in the wq_has_sleeper().
* See the comment above prepare_to_wait(), we need to
* ensure that subsequent tests in this thread can't be
* reordered with __add_wait_queue() in _qproc() paths.
*/
smp_mb();
}
}
/*