Re: [Patch v3 3/6] cpufreq: qcom-cpufreq-hw: Add dcvs interrupt support

[Thara Gopinath]

On 7/12/21 12:41 AM, Viresh Kumar wrote:
> On 09-07-21, 11:37, Thara Gopinath wrote:
> > On 7/9/21 2:46 AM, Viresh Kumar wrote:
> > > > @@ -389,6 +503,10 @@ static int qcom_cpufreq_hw_cpu_exit(struct cpufreq_policy *policy)
> > > >    	dev_pm_opp_remove_all_dynamic(cpu_dev);
> > > >    	dev_pm_opp_of_cpumask_remove_table(policy->related_cpus);
> > > > +	if (data->lmh_dcvs_irq > 0) {
> > > > +		devm_free_irq(cpu_dev, data->lmh_dcvs_irq, data);
> > >
> > > Why using devm variants here and while requesting the irq ?
>
> Missed this one ?

Yep. I just replied to Bjorn's email on this. I will move to non devm 
version.

> > > > +		cancel_delayed_work_sync(&data->lmh_dcvs_poll_work);
> > > > +	}
> > >
> > > Please move this to qcom_cpufreq_hw_lmh_exit() or something.
> >
> > Ok.
> >
> > > Now with sequence of disabling interrupt, etc, I see a potential
> > > problem.
> > >
> > > CPU0                                    CPU1
> > >
> > > qcom_cpufreq_hw_cpu_exit()
> > > -> devm_free_irq();
> > >                                           qcom_lmh_dcvs_poll()
> > >                                           -> qcom_lmh_dcvs_notify()
> > >                                             -> enable_irq()
> > >
> > > -> cancel_delayed_work_sync();
> > >
> > >
> > > What will happen if enable_irq() gets called after freeing the irq ?
> > > Not sure, but it looks like you will hit this then from manage.c:
> > >
> > > WARN(!desc->irq_data.chip, KERN_ERR "enable_irq before
> > >                                        setup/request_irq: irq %u\n", irq))
> > >
> > > ?
> > >
> > > You got a chicken n egg problem :)
> >
> > Yes indeed! But also it is a very rare chicken and egg problem.
> > The scenario here is that the cpus are busy and running load causing a
> > thermal overrun and lmh is engaged. At the same time for this issue to be
> > hit the cpu is trying to exit/disable cpufreq.
>
> Yes, it is a very specific case but it needs to be resolved anyway. You don't
> want to get this ever :)
>
> > Calling
> > cancel_delayed_work_sync first could solve this issue, right ?
> > cancel_delayed_work_sync guarantees the work not to be pending even if
> > it requeues itself on return. So once the delayed work is cancelled, the
> > interrupts can be safely disabled. Thoughts ?
>
> I don't think even that would provide such guarantees to you here, as there is
> a chance the work gets queued again because of an interrupt that triggers right
> after you cancel the work.
>
> The basic way of solving such issues is that once you cancel something, you need
> to guarantee that it doesn't get triggered again, no matter what.
>
> The problem here I see is with your design itself, both delayed work and irq can
> enable each other, so no matter which one you disable first, won't be
> sufficient. You need to fix that design somehow.

So I really need the interrupt to fire and then the timer to kick in and 
take up the monitoring. I can think of introducing a variable 
is_disabled which is updated and read under a spinlock. 
qcom_cpufreq_hw_cpu_exit can hold the spinlock and set is_disabled to 
true prior to cancelling the work queue or disabling the interrupt. 
Before re-enabling the interrupt or re-queuing the work in 
qcom_lmh_dcvs_notify, is_disabled can be read and checked.

But does this problem not exist in target_index , fast_switch etc also ? 
One cpu can be disabling and the other one can be updating the target 
right?

>

--
Warm Regards
Thara (She/Her/Hers)