Re: [PATCH] fix refrigerator() vs thaw_process() race

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

On 02/19, Pavel Machek wrote:
> 
> > refrigerator() can miss a wakeup, "wait event" loop needs a proper memory
> > ordering.
> > 
> > Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
> > 
> > --- WQ/kernel/power/process.c~WAKE	2007-02-18 22:56:49.000000000 +0300
> > +++ WQ/kernel/power/process.c	2007-02-19 01:04:26.000000000 +0300
> > @@ -46,8 +46,10 @@ void refrigerator(void)
> >  	recalc_sigpending(); /* We sent fake signal, clean it up */
> >  	spin_unlock_irq(&current->sighand->siglock);
> >  
> > -		current->state = TASK_UNINTERRUPTIBLE;
> > +		set_current_state(TASK_UNINTERRUPTIBLE);
> 
> 
> Looks okay to me... but this one liner would be exactly as effective,
> right?

I think no, with this one liner we have

	while (frozen(current)) {
		// ------ WINDOW ------------
		set_current_state(TASK_UNINTERRUPTIBLE);
		schedule();
	}

What if thaw_process() happens in the window above?

We need the barrier exactly because LOAD (check condition) should not
come before STORE (set task->state).

Oleg.