Re: [kernel.org users] [KORG] Panics on master backend

[Yong Zhang <yong.zhang0@gmail.com>]

On Fri, Aug 26, 2011 at 03:57:39PM +0200, Oleg Nesterov wrote:
> On 08/26, Yong Zhang wrote:
> >
> > On Thu, Aug 25, 2011 at 03:54:29PM +0200, Oleg Nesterov wrote:
> > >
> > > Of course it is not TASK_RUNNING, but it can be running or not.
> >
> > Yup. Before we go beyond ttwu_remote() in ttwu(), 'cpu' is not safe.
> > For example, wait_event() could be preempted in between.
> >
> > But after we go beyond ttwu_remote(), ->pi_lock will stabilize it.
> 
> Yes.
> 
> > So after we take Oleg's suggestion("task_cpu(p) == smp_processor_id()"),
> > things we left is just how to account stat correctly.
> 
> Imho, we don't really care. This race is very unlikely, 

Yup.

> and I think
> that the "wrong" cpu argument in ttwu_stat() is harmless.

Hmm, the affected accounting is sched_domain->ttwu_wake_remote.

> > @@ -2696,7 +2697,12 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
> >  	success = 1; /* we're going to change ->state */
> >  	cpu = task_cpu(p);
> >  
> > -	if (p->on_rq && ttwu_remote(p, wake_flags))
> > +	/*
> > +	 * read cpu for another time if ttwu_remote() success,
> > +	 * just to prevent task migration in between, otherwise
> > +	 * we maybe account stat incorrectly.
> > +	 */
> > +	if (p->on_rq && ttwu_remote(p, wake_flags, &cpu))
> 
> I don't think this makes the things better. p->on_rq can be already
> false or ttwu_remote() can fail, in this case we still use the result
> of initial task_cpu().

Ah, My code (ttwu_remote(p, wake_flags, &cpu)) just catch the case:
task is dancing with 'p->on_rq == true'.

But forget the case: task has danced with 'p->on_rq == false' in the end.

So we should reread task_cpu(p) if 'p->on_cpu == true &&
__ARCH_WANT_INTERRUPTS_ON_CTXSW == y', but the code will be a little ugly
and it only aims to account stat more correctly.
We need some balance here :)
BTW, I don't think we should care much on the 'incorrect stat' either
if we don't make intolerable mistake.

Thanks,
Yong