On Sun, May 08, 2016 at 09:56:10PM -0700, Davidlohr Bueso wrote:
> Readers that are awoken will expect a nil ->task indicating
> that a wakeup has occurred. There is a mismatch between the
> smp_mb() and its documentation, in that the serialization is
> done between reading the task and the nil store. Furthermore,
> in addition to having the overlapping of loads and stores to
> waiter->task guaranteed to be ordered within that CPU, both
> wake_up_process() originally and now wake_q_add() already
> imply barriers upon successful calls, which serves the comment.
>
> Just atomically do a xchg() and simplify the whole thing. We can
> use relaxed semantics as before mentioned in addition to the
> barrier provided by wake_q_add(), delaying there is no risk in
> reordering with the actual wakeup.
> @@ -190,24 +189,18 @@ __rwsem_mark_wake(struct rw_semaphore *sem,
> next = sem->wait_list.next;
> loop = woken;
> do {
> + struct task_struct *tsk;
> +
> waiter = list_entry(next, struct rwsem_waiter, list);
> next = waiter->list.next;
> - tsk = waiter->task;
> - /*
> - * Make sure we do not wakeup the next reader before
> - * setting the nil condition to grant the next reader;
> - * otherwise we could miss the wakeup on the other
> - * side and end up sleeping again. See the pairing
> - * in rwsem_down_read_failed().
> - */
> - smp_mb();
> - waiter->task = NULL;
> +
> + tsk = xchg_relaxed(&waiter->task, NULL);
> wake_q_add(wake_q, tsk);
Not a great fan of this patch; it again doesn't fix the race, and
smp_store_release() is a cheaper option on x86.
