Re: [PATCH v7] sched: Consolidate cpufreq updates

[Qais Yousef]

On 10/11/24 10:34, Christian Loehle wrote:

> > @@ -796,7 +802,8 @@ int __sched_setscheduler(struct task_struct *p,
> >  		__setscheduler_params(p, attr);
> >  		__setscheduler_prio(p, newprio);
> >  	}
> > -	__setscheduler_uclamp(p, attr);
> > +
> > +	update_cpufreq = __setscheduler_uclamp(p, attr);
> >  
> >  	if (queued) {
> >  		/*
> > @@ -811,7 +818,18 @@ int __sched_setscheduler(struct task_struct *p,
> >  	if (running)
> >  		set_next_task(rq, p);
> >  
> > -	check_class_changed(rq, p, prev_class, oldprio);
> > +	update_cpufreq |= check_class_changed(rq, p, prev_class, oldprio);
> > +
> > +	/*
> > +	 * Changing class or uclamp value implies requiring to send cpufreq
> > +	 * update.
> > +	 */
> > +	if (update_cpufreq) {
> > +		if (running)
> > +			update_cpufreq_current(rq);
> > +		else if (queued)
> > +			cpufreq_update_util(rq, SCHED_CPUFREQ_TASK_ENQUEUED);
> > +	}
> 
> cpufreq_update_util() -> sugov_should_update_freq() -> cpufreq_this_cpu_can_update()
> relies on smp_processor_id(), should this move this below the
> preempt_disable() to avoid sending an update from an illegal CPU?

We are holding the task_rq_lock() with interrupts disabled, we should be okay
no?

Did you see an actual warning from LOCKDEP?