On Thu, 22 Aug 2024 at 06:28, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
>
> wakeup_var() takes the wait queue head spinlock
That's the part you missed.
wake_up_var() does no such thing. It calles __wake_up_bit(), which does this:
if (waitqueue_active(wq_head))
__wake_up(wq_head, TASK_NORMAL, 1, &key);
and that waitqueue_active() is lockless.
That's the thing that wants the barrier (or something else that
guarantees that the waiters actually *see* the new state after they've
added themselves to the wait-list, and before we then look at the
wait-list being empty).
See the comment in <linux/wait.h> above the waitqueue_active().
And yes, this is annoying and subtle and has caused bugs. But taking
waitqueue locks just to notice nobody is waiting is very expensive
indeed, and much more expensive than a memory barrier when the common
case is typically often "nobody is waiting".
Linus
