[PATCH] drivers: perf: arm: implement CPU_PM notifier

[Lorenzo Pieralisi]

On Wed, Feb 24, 2016 at 02:53:02PM -0500, Ashwin Chaugule wrote:
> Hi Lorenzo,
> 
> On 24 February 2016 at 12:35, Lorenzo Pieralisi
> <lorenzo.pieralisi at arm.com> wrote:
> > On Wed, Feb 24, 2016 at 09:20:22AM -0700, Mathieu Poirier wrote:
> >> On 23 February 2016 at 11:22, Lorenzo Pieralisi
> >
> > [...]
> >
> >> > +static int cpu_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
> >> > +                            void *v)
> >> > +{
> >> > +       struct arm_pmu *armpmu = container_of(b, struct arm_pmu, cpu_pm_nb);
> >> > +       struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
> >> > +       int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
> >> > +
> >> > +       if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
> >> > +               return NOTIFY_DONE;
> >> > +
> >> > +       /*
> >> > +        * Always reset the PMU registers on power-up even if
> >> > +        * there are no events running.
> >> > +        */
> >> > +       if (cmd == CPU_PM_EXIT && armpmu->reset)
> >> > +               armpmu->reset(armpmu);
> >>
> >> I think this patch does the right thing but I can't get the above
> >> reset.  Wouldn't it be better to do the reset as part of the
> >> CPU_PM_EXIT case below?  At this point nothing tells us the CPU won't
> >> go back down before the event is enabled, wasting the cycle needed to
> >> reset the PMU.
> >
> > The logic goes, if the cpu is woken up and it has no events enabled,
> > if we do not reset it (mind, ->reset here sets the PMU register values
> > to a sane default, some of them are architecturally UNKNOWN on reset, it
> > does NOT reset the PMU) _and_ we subsequently install an event on it we
> > do have a problem, that's why whenever a core gets out of low-power we
> > have to reset its pmu.
> 
> Dont we blow out the previous value in the counter when installing an
> event? Or has that changed lately? IIRC, there was some initial value
> we'd program into the counter to avoid missing the overflow event.
> (sorry its been too long) ;)

If you mean there is no need of resetting the value since we are
overwriting it anyway you should see ->reset from a PMU unit POW
not just an event. If you look at the ->reset method for eg v8, you
will see that the reset method operates on all counters and the PMU
unit as a whole, that's the only sane way of setting up the PMU unit
before enabling single counters (some registers are UNKNOWN at reset).

To make my reply to Mathieu clearer, the ->reset method contains
operations (eg writing PMCR counters reset bits) that do carry out
counters reset, what I wanted to say is that the ->reset method
does not by itself drive the PMU HW reset signal, that's what the
power controller does when it resets the CPU on power up.

The PMU ->reset method must be called on a cpu on power-up, to make
sure PMU HW is set-up to sane default values and ready to be used (ie
install counters), for instance on v8 all counters must be disabled
(irq inclusive) and reset, that's what armv8pmu_reset() does.

I hope this makes things clearer.

Thanks,
Lorenzo