Re: [PATCH 16/19] perf: Introduce positive capability for sampling

On 2025-08-26 2:11 pm, Leo Yan wrote:
> On Wed, Aug 13, 2025 at 06:01:08PM +0100, Robin Murphy wrote:
>> Sampling is inherently a feature for CPU PMUs, given that the thing
>> to be sampled is a CPU context. These days, we have many more
>> uncore/system PMUs than CPU PMUs, so it no longer makes much sense to
>> assume sampling support by default and force the ever-growing majority
>> of drivers to opt out of it (or erroneously fail to). Instead, let's
>> introduce a positive opt-in capability that's more obvious and easier to
>> maintain.
> 
> [...]
> 
>> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
>> index 369e77ad5f13..dbd52851f5c6 100644
>> --- a/drivers/perf/arm_spe_pmu.c
>> +++ b/drivers/perf/arm_spe_pmu.c
>> @@ -955,7 +955,8 @@ static int arm_spe_pmu_perf_init(struct arm_spe_pmu *spe_pmu)
>>   	spe_pmu->pmu = (struct pmu) {
>>   		.module = THIS_MODULE,
>>   		.parent		= &spe_pmu->pdev->dev,
>> -		.capabilities	= PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
>> +		.capabilities	= PERF_PMU_CAP_SAMPLING |
>> +				  PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
>>   		.attr_groups	= arm_spe_pmu_attr_groups,
>>   		/*
>>   		 * We hitch a ride on the software context here, so that
> 
> The change in Arm SPE driver looks good to me.
> 
> I noticed you did not set the flag for other AUX events, like Arm
> CoreSight, Intel PT and bts. The drivers locate in:
> 
>    drivers/hwtracing/coresight/coresight-etm-perf.c
>    arch/x86/events/intel/bts.c
>    arch/x86/events/intel/pt.c
> 
> Genearlly, AUX events generate interrupts based on AUX ring buffer
> watermark but not the period. Seems to me, it is correct to set the
> PERF_PMU_CAP_SAMPLING flag for them.

This cap is given to drivers which handle event->attr.sample_period and 
call perf_event_overflow() - or in a few rare cases, 
perf_output_sample() directly - to do something meaningful with it, 
since the intent is to convey "I properly handle events for which 
is_sampling_event() is true". My understanding is that aux events are 
something else entirely, but I'm happy to be corrected.

Otherwise, perhaps this suggests it deserves to be named a little more 
specifically for clarity, maybe PERF_CAP_SAMPLING_EVENTS?

Thanks,
Robin.

> A special case is Arm CoreSight legacy sinks (like ETR/ETB, etc)
> don't has interrupt. We might need set or clear the flag on the fly
> based on sink type:
> 
> diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.c b/drivers/hwtracing/coresight/coresight-etm-perf.c
> index f1551c08ecb2..404edc94c198 100644
> --- a/drivers/hwtracing/coresight/coresight-etm-perf.c
> +++ b/drivers/hwtracing/coresight/coresight-etm-perf.c
> @@ -433,6 +433,11 @@ static void *etm_setup_aux(struct perf_event *event, void **pages,
>          if (!sink)
>                  goto err;
>   
> +       if (coresight_is_percpu_sink(sink))
> +               event->pmu.capabilities = PERF_PMU_CAP_SAMPLING;
> +       else
> +               event->pmu.capabilities &= ~PERF_PMU_CAP_SAMPLING;
> +
> 
> Thanks,
> Leo