Re: [PATCH v6 2/2] sched: update the rq->avg_idle when a task is moved to an idle CPU

[Shijie Huang]

On 16/12/2025 16:47, Vincent Guittot wrote:
> On Tue, 16 Dec 2025 at 08:39, Shijie Huang
> <shijie@xxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> > On 16/12/2025 15:17, Vincent Guittot wrote:
> > > On Tue, 16 Dec 2025 at 07:22, Shijie Huang
> > > <shijie@xxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > > On 13/12/2025 09:36, Vincent Guittot wrote:
> > > > > put_prev_task_idle() would be a better place to call
> > > > > update_rq_avg_idle() because this is when we leave idle.
> > > > The update_rq_avg_idle() is not only called by current CPU, but also
> > > > called by
> > > >
> > > > other CPUs. For example, the try_to_wake_up(), update_rq_avg_idle() is
> > > > called by
> > > >
> > > > the other CPUs. So enqueue_task() is a good place.
> > > But put_prev_task_idle() is called by local CPU whenever it leaves
> > > idle so instead of trying to catch all places that could make the CPU
> > > leave idle it's better to use this single place.
> > > And as you mentioned, put_prev_task_idle is only called by local CPU
> > > whereas enqueue_task can be called by all CPUs creating useless
> > > pressure in the variable.
> > The rq->idle_stamp is set at sched_balance_newidle().  then we call
> > update_rq_avg_idle()
> >
> > in put_prev_task_idle() right now. How can we update the rq->avg_idle?
> I'm not sure I understand your point.
>
> rq->avg_idle tracks idle time. The easiest way would be to use
> - set_next_task_idle() when we enter idle
> - put_prev_task_idle() when we exit idle
>
> Except that sched_balance_newidle() can be long and the time should be
> accounted as idle time too. So instead of using set_next_task_idle(),
> we use sched_balance_newidle() to set . Which is okay because
> sched_balance_newidle() is always called before going to idle.

Thanks for the explanations.

It seems that put_prev_task_idle() is really a better place to call
update_rq_avg_idle(). Let me think it for a while :)


Thanks
Huang Shijie



>