Re: futex code and barriers

[Oleg Nesterov <oleg@tv-sign.ru>]

On 04/29, Peter Zijlstra wrote:
> Hi All,
> 
> While looking through the futex code I stumbled upon the following bit:
> 
> kernel/futex.c:
> 
> 	/* add_wait_queue is the barrier after __set_current_state. */

As for me, the comment is very confusing at least.

> 	__set_current_state(TASK_INTERRUPTIBLE);
> 	add_wait_queue(&q.waiters, &wait);

Not sure I understand this code, but probably it is correct.

Yes, add_wait_queue() is not a barrier, and both __set_current_state()
and the "!plist_node_empty()" check below can leak into the
add_wait_queue's critical section.

But wake_futex()->wake_up_all() has to lock/unlock the same q->lock,
so I think we can't miss the event.

IOW, when wake_futex()->wake_up_all() takes q->lock, it must see
TASK_INTERRUPTIBLE.

If wake_futex() takes q->lock before us, we must see the result
of plist_del() after add_wait_queue() (more precisely, after
add_wait_queue()->spin_lock(q->lock).

But I'd like to know maintainer's opinion, I don't trust myself ;)

Oleg.