Re: [patch] sys_epoll ...

[Andrew Morton <akpm@digeo.com>]

Davide Libenzi wrote:
> 
> On Sun, 20 Oct 2002, Andrew Morton wrote:
> 
> > +                       if (ep->eventcnt || !timeout)
> > +                               break;
> > +                       if (signal_pending(current)) {
> > +                               res = -EINTR;
> > +                               break;
> > +                       }
> > +
> > +                       set_current_state(TASK_INTERRUPTIBLE);
> > +
> > +                       write_unlock_irqrestore(&ep->lock, flags);
> > +                       timeout = schedule_timeout(timeout);
> >
> > You should set current->state before performing the checks.
> 
> Why this Andrew ?
> 

Well I'm assuming that you don't want to sleep if, say,
ep->eventcnt is non-zero.  The code is currently (simplified):

	add_wait_queue(...);
	if (ep->eventcnt)
		break;
	/* window here */
	set_current_state(TASK_INTERRUPTIBLE);
	schedule();

If another CPU increments eventcnt and sends this task a wakeup in that
window, it is missed and we still sleep.  The conventional fix for that
is:

	add_wait_queue(...);
	set_current_state(TASK_INTERRUPTIBLE);
	if (ep->eventcnt)
		break;
	/* harmless window here */
	schedule();

So if someone delivers a wakeup in the "harmless window" then this task
still calls schedule(), but the wakeup has turned the state from
TASK_INTERRUPTIBLE into TASK_RUNNING, so the schedule() doesn't actually
take this task off the runqueue.  This task will zoom straight through the
schedule() and will then loop back and notice the incremented ep->eventcnt.

So it is important that the waker increment eventcnt _before_ delivering
the wake_up, too.