Re: [RFC PATCH 5/6] sched/fair: Select an energy-efficient CPU on task wake-up

[Quentin Perret]

On Friday 23 Mar 2018 at 16:00:59 (+0000), Morten Rasmussen wrote:
> On Thu, Mar 22, 2018 at 09:27:43AM -0700, Joel Fernandes wrote:
> > Hi,
> > 
> > On Tue, Mar 20, 2018 at 2:43 AM, Dietmar Eggemann
> > <dietmar.eggemann@xxxxxxx> wrote:

[...]

> > Is it possible that before the wakeup, the task's affinity is changed
> > so that p->cpus_allowed no longer contains prev_cpu ? In that case
> > prev_energy wouldn't matter since previous CPU is no longer an option?
> 
> It is possible to wake-up with a disallowed prev_cpu. In fact
> select_idle_sibling() may happily return a disallowed cpu in that case.
> The mistake gets fixed in select_task_rq() which uses
> select_fallback_rq() to find an allowed cpu instead.
> 
> Could we fix the issue in find_energy_efficient_cpu() by a simple test
> like below
> 
> if (cpumask_test_cpu(prev_cpu, &p->cpus_allowed))
> 	prev_energy = best_energy = compute_energy(p, prev_cpu);
> else
> 	prev_energy = best_energy = ULONG_MAX;

Right, that should work. I'll change this in v2.

Thanks,
Quentin