Re: [PATCH 4/9] perf/x86/intel: Support hardware TopDown metrics

[Peter Zijlstra <peterz@infradead.org>]

On Tue, May 21, 2019 at 02:40:50PM -0700, kan.liang@linux.intel.com wrote:
> The 8bit metrics ratio values lose precision when the measurement period
> gets longer.
> 
> To avoid this we always reset the metric value when reading, as we
> already accumulate the count in the perf count value.
> 
> For a long period read, low precision is acceptable.
> For a short period read, the register will be reset often enough that it
> is not a problem.

> The PERF_METRICS may report wrong value if its delta was less than 1/255
> of SLOTS (Fixed counter 3).
> 
> To avoid this, the PERF_METRICS and SLOTS registers have to be reset
> simultaneously. The slots value has to be cached as well.

That doesn't sound like it is NMI-safe.



> RDPMC
> =========
> The TopDown can be collected per thread/process. To use TopDown
> through RDPMC in applications on Icelake, the metrics and slots values
> have to be saved/restored during context switching.
> 
> Add specific set_period() to specially handle the slots and metrics
> event. Because,
>  - The initial value must be 0.
>  - Only need to restore the value in context switch. For other cases,
>    the counters have been cleared after read.

So the above claims to explain RDPMC, but doesn't mention that magic
value below at all. In fact, I don't see how the above relates to RDPMC
at all.

> @@ -2141,7 +2157,9 @@ static int x86_pmu_event_idx(struct perf_event *event)
>  	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
>  		return 0;
>  
> -	if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
> +	if (is_metric_idx(idx))
> +		idx = 1 << 29;

I can't find this in the SDM RDPMC description. What does it return?

> +	else if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
>  		idx -= INTEL_PMC_IDX_FIXED;
>  		idx |= 1 << 30;
>  	}