Re: [PATCH] It may not be assumed that wake_up(), finish_wait() and co. imply a memory barrier

[David Howells <dhowells@redhat.com>]

Linus Torvalds <torvalds@linux-foundation.org> wrote:

> I think we _have_ to imply a smp_wmb() in the wakup semantics, because 
> otherwise sleepers can't do anything sane (no amount of barriers on the 
> sleeping side will help). IOW, there basically has to be an implied write 
> barrier between the thing that causes an event to become true, and the 
> thing that turns 'task->state' back to RUNNING.

Well, Ingo's point is it could be left up to the caller of wake_up() to supply
the barrier:

	*my_variable = 1234;
	smp_wmb();
	wake_up(&my_queue);

or:

	spin_lock(&my_lock);
	*my_variable = 1234;
	wake_up(&my_queue);
	spin_unlock(&my_lock);

on the condition that the sleeper also gets the lock.

Also, he points out that wake_up() and co. may not insert useful memory barrier
at all, for three reasons:

 (1) If there's no-one to wake up, then certain wake functions will return
     immediately.

 (2) If there's no-one to wake up, then other wake functions will only impose
     LOCK and UNLOCK barriers.

 (3) If someone supplies a special awakener instead of default_wake_function(),
     then they can bypass try_to_wake_up() and whatever barrier that implies.

     Though as far as I can see, if you want to wake someone up, you *have* to
     go through try_to_wake_up().

What I'd like to say is that wake_up() and friends _will_ interpose at least a
write barrier _if_ they wake anything up (which is more or less what you said
above).  If they don't wake anything up, then there's no need for a memory
barrier between the assignment to my_variable and the non-existent alterations
to the state of the task not being awoken.

David