[Intel-gfx] [PATCH 01/13] drm/i915: Add a sw fence for collecting up dma fences

[Chris Wilson]

On Thu, Aug 25, 2016 at 02:49:59PM +0300, Joonas Lahtinen wrote:
> > +	 * we move the task_list from this the next ready fence to the tail of
> > +	 * the original fence's task_list (and so added to the list to be
> > +	 * woken).
> > +	 */
> > +	smp_mb__before_spinlock();
> > +	if (!list_empty_careful(&x->task_list)) {
> 
> if (list_empty_careful()
> 	return;

It's just broken. I added it recently after reading

void finish_wait(wait_queue_head_t *q, wait_queue_t *wait)
{
        unsigned long flags;

        __set_current_state(TASK_RUNNING);
        /*
         * We can check for list emptiness outside the lock
         * IFF:
         *  - we use the "careful" check that verifies both
         *    the next and prev pointers, so that there cannot
         *    be any half-pending updates in progress on other
         *    CPU's that we haven't seen yet (and that might
         *    still change the stack area.
         * and
         *  - all other users take the lock (ie we can only
         *    have _one_ other CPU that looks at or modifies
         *    the list).
         */
        if (!list_empty_careful(&wait->task_list)) {
                spin_lock_irqsave(&q->lock, flags);
                list_del_init(&wait->task_list);
                spin_unlock_irqrestore(&q->lock, flags);
        }
}

and convinced myself that it was also safe to apply here.
Turns out that spinlock is very hard to avoid.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre