question about blk_schedule_flush_plug

question about blk_schedule_flush_plug

[Con Kolivas]

I was looking at the scheduler changes going into 2.6.39 again and wondered 
about the use of blk_schedule_flush_plug smack in the middle of schedule()

It looks like this:
			if (blk_needs_flush_plug(prev)) {
				raw_spin_unlock(&rq->lock);
				blk_schedule_flush_plug(prev);
				raw_spin_lock(&rq->lock);
			}

Now call me suspicious but to my eyes it looks really dubious unlocking the 
runqueue like that right in the heart of schedule().

Comments?

Thanks,
Con
-- 
-ck

Re: question about blk_schedule_flush_plug

[Peter Zijlstra]

On Mon, 2011-05-23 at 17:05 +1000, Con Kolivas wrote:
> I was looking at the scheduler changes going into 2.6.39 again and wondered 
> about the use of blk_schedule_flush_plug smack in the middle of schedule()
> 
> It looks like this:
> 			if (blk_needs_flush_plug(prev)) {
> 				raw_spin_unlock(&rq->lock);
> 				blk_schedule_flush_plug(prev);
> 				raw_spin_lock(&rq->lock);
> 			}
> 
> Now call me suspicious but to my eyes it looks really dubious unlocking the 
> runqueue like that right in the heart of schedule().
> 
> Comments?

Yeah, that's quite all-right, both pre_schedule() and idle_balance()
already did that too.

We can safely release the runqueue lock after deactivate_task() and
before put_prev_task().

Re: question about blk_schedule_flush_plug

[Mike Galbraith]

On Mon, 2011-05-23 at 17:05 +1000, Con Kolivas wrote:
> I was looking at the scheduler changes going into 2.6.39 again and wondered 
> about the use of blk_schedule_flush_plug smack in the middle of schedule()
> 
> It looks like this:
> 			if (blk_needs_flush_plug(prev)) {
> 				raw_spin_unlock(&rq->lock);
> 				blk_schedule_flush_plug(prev);
> 				raw_spin_lock(&rq->lock);
> 			}
> 
> Now call me suspicious but to my eyes it looks really dubious unlocking the 
> runqueue like that right in the heart of schedule().
> 
> Comments?

Releasing/retaking rq->lock is nothing new:

static void idle_balance(int this_cpu, struct rq *this_rq)
{
...
        /*
         * Drop the rq->lock, but keep IRQ/preempt disabled.
         */
        raw_spin_unlock(&this_rq->lock);

See also need_resched, and double_lock_balance() instances.

	-Mike