Re: [PATCH 2/4] sched/deadline: Improve admission control for asymmetric CPU capacities

[Qais Yousef]

On 04/14/20 15:29, Valentin Schneider wrote:
> 
> On 14/04/20 12:40, Qais Yousef wrote:
> >
> > I haven't followed this discussion closely, so I could be missing something
> > here.
> >
> > In sched_cpu_dying() we call set_rq_offline() which clears the cpu in
> > rq->rd->online.
> >
> > So the way I read the code
> >
> >       rd->online = cpumask_and(rd->span, cpu_active_mask)
> >
> > But I could have easily missed some detail.
> >
> 
> sched_cpu_dying() is wayyyy below sched_cpu_deactivate(). This doesn't help
> at all for the dl_cpu_busy() check in sched_cpu_deactivate().

It'd makes more sense to introduce rd->active then rather than a PERCPU global
variable IMHO.

> 
> > Regardless, it seems to me that DL is working around something not right in the
> > definition of rd->span or using the wrong variable.
> >
> 
> What DL is doing now is fine, it only needs to be aligned with the active
> mask (which it is). We're making things a bit trickier by adding capacity
> values into the mix.

I'm still worried that others might be affected by a similar problem.

Will dig deeper next week when I'm back from hols.

Thanks

--
Qais Yousef