Re: [PATCH 2/6] move_task_off_dead_cpu: take rq->lock around select_fallback_rq()

[Peter Zijlstra <peterz@infradead.org>]

On Wed, 2010-03-24 at 17:33 +0100, Oleg Nesterov wrote:
> On 03/24, Peter Zijlstra wrote:
> >
> > Yeah, you made a few good points in 0/6, am now staring at the code on
> > how to close those holes, hope to post something sensible soon.
> 
> Yes, great.
> 
> Speaking of 0/6, I forgot to ask a couple more question...
> 
> try_to_wake_up() does task_rq_lock() which checks TASK_WAKING. Perhaps
> it shouldn't ? I mean, perhaps try_to_wake_up() can take rq->lock without
> checking task->state. It can never race with the owner of TASK_WAKING,
> before anything else we check "p->state & state".

You're right, but creating a special task_rq_lock() for ttwu() went a
little far, but now that we can remove all that again, this too should
be good again.


> And a stupid question. While doing these changes I was really, really
> puzzled by task_rq_lock() which does
> 
> 	local_irq_save(*flags);
> 	rq = task_rq(p);
> 	raw_spin_lock(&rq->lock);
> 
> to the point, I even tried to read the comment which says:
> 
> 	Note the ordering: we can safely lookup the task_rq without
> 	explicitly disabling preemption.
> 
> Could you please explain what does this mean? IOW, why can't we do
> 
> 	rq = task_rq(p);
> 	raw_spin_lock_irqsave(&rq->lock, flags);
> 
> instead?

I'm not sure why that is the case, v2.6.14:kernel/sched.c already has
that. Ingo can you remember any reason for this or should we change the
code like Oleg suggests?