Re: [PATCH linux-next] sched: Fix broken setscheduler()

[Juri Lelli <juri.lelli@gmail.com>]

On Thu, 6 Mar 2014 12:58:25 +0100
Peter Zijlstra <peterz@infradead.org> wrote:

> On Wed, Mar 05, 2014 at 11:29:31PM -0500, Steven Rostedt wrote:
> > I decided to run my tests on linux-next, and my wakeup_rt tracer was
> > broken. After running a bisect, I found that the problem commit was:
> > 
> >    linux-next commit c365c292d059
> >    "sched: Consider pi boosting in setscheduler()"
> > 
> > And the reason the wake_rt tracer test was failing, was because it had
> > no RT task to trace. I first noticed this when running with
> > sched_switch event and saw that my RT task still had normal SCHED_OTHER
> > priority. Looking at the problem commit, I found:
> > 
> >  -       p->normal_prio = normal_prio(p);
> >  -       p->prio = rt_mutex_getprio(p);
> > 
> > With no
> > 
> >  +       p->normal_prio = normal_prio(p);
> >  +       p->prio = rt_mutex_getprio(p);
> > 
> > Reading what the commit is suppose to do, I realize that the p->prio
> > can't be set if the task is boosted with a higher prio, but the
> > p->normal_prio still needs to be set regardless, otherwise, when the
> > task is deboosted, it wont get the new priority.
> > 
> > The p->prio has to be set before "check_class_changed()" is called,
> > otherwise the class wont be changed.
> 
> So Juri had a different patch for this problem:
> 
>   http://lkml.kernel.org/r/20140301191838.d15d03112b2598a671dac22c@gmail.com
> 
> > Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
> > ---
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 4600bca..b1cc871 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -3198,6 +3198,7 @@ static void __setscheduler_params(struct task_struct *p,
> >  	 * getparam()/getattr() don't report silly values for !rt tasks.
> >  	 */
> >  	p->rt_priority = attr->sched_priority;
> > +	p->normal_prio = normal_prio(p);
> >  	set_load_weight(p);
> >  }
> 
> Now; if I'm reading things right, normal_prio is the unboosted priority
> of a task. And we should indeed keep setting that, otherwise the unboost
> doesn't know where it should go.
> 
> Juri put that in __setscheduler(), but I think that's wrong because the
> rt_mutex_check_prio() case in __sched_setscheduler() still needs to
> update this.
> 

Oh, right. Missed that.

> > @@ -3207,6 +3208,8 @@ static void __setscheduler(struct rq *rq, struct task_struct *p,
> >  {
> >  	__setscheduler_params(p, attr);
> >  
> > +	p->prio = rt_mutex_getprio(p);
> > +
> >  	if (dl_prio(p->prio))
> >  		p->sched_class = &dl_sched_class;
> >  	else if (rt_prio(p->prio))
> > 
> 
> And when we call this we're sure to not be boosted; so this is
> effectively the same as Juri has:
> 
>   p->prio = p->normal_prio = normal_prio(p)
> 
> Seeing how rt_mutex_getprio() and normal_prio() are the same under these
> conditions.
> 
> 

Yes. I think you can go with

 p->prio = p->normal_prio

and save a few checks in rt_mutex_getprio().

Thanks,

- Juri