Re: [PATCH 6/6] arm64: perf: Add support for chaining counters

[Suzuki K Poulose]

On 18/05/18 15:57, Robin Murphy wrote:
> One more thing now that I've actually looked at the Arm ARM...
>
> On 18/05/18 11:22, Suzuki K Poulose wrote:
> [...]
> > +static inline void armv8pmu_write_event_type(struct perf_event *event)
> > +{
> > +ÂÂÂ struct hw_perf_event *hwc = &event->hw;
> > +ÂÂÂ int idx = hwc->idx;
> > +
> > +ÂÂÂ /*
> > +ÂÂÂÂ * For chained events, write the high counter event type
> > +ÂÂÂÂ * followed by the low counter.
> > +ÂÂÂÂ */
> > +ÂÂÂ if (armv8pmu_event_is_chained(event)) {
> > +ÂÂÂÂÂÂÂ u32 chain_evt = ARMV8_PMUV3_PERFCTR_CHAIN;
> > +
> > +ÂÂÂÂÂÂÂ /* Set the filters as that of the main event for chain */
> > +ÂÂÂÂÂÂÂ chain_evt |= hwc->config_base & ~ARMV8_PMU_EVTYPE_EVENT;
>
> The description of the chain event says that the filtering must only be set on the lower counter, and that the chain event itself should be set to count everything.

You're right. I intended to fix it, but missed in the rebases.
Thanks for pointing out.

> > +ÂÂÂÂÂÂÂ armv8pmu_write_evtype(idx, chain_evt);
> > +ÂÂÂÂÂÂÂ isb();
> > +ÂÂÂÂÂÂÂ idx--;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ armv8pmu_write_evtype(idx, hwc->config_base);
>
> It also says that the 'real' event should be set up first and the chain event second, with the rather ominous warning of "If software does not program the event in this way, the count becomes UNPREDICTABLE."

Yep, will fix it.

Thanks for the review

Suzuki