Re: [PATCH v6 1/8] perf/core: Allow multiple AUX PMU events with the same module

From: Adrian Hunter
Date: Thu Sep 05 2024 - 03:13:21 EST


On 4/09/24 22:35, Leo Yan wrote:
> On 9/3/2024 11:06 AM, Adrian Hunter wrote:
>>> @@ -12345,9 +12345,16 @@ perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
>>>
>>> /*
>>> * If both events generate aux data, they must be on the same PMU
>>> + * module but can be with different PMU instances.
>>> + *
>>> + * For a built-in PMU module, the 'pmu->module' pointer is NULL,
>>> + * thus it is not feasible to compare the module pointers when
>>> + * AUX PMU drivers are built into the kernel image. Instead,
>>> + * comparing the .setup_aux() callback pointer can determine if
>>> + * the two PMU events come from the same PMU driver.
>>> */
>>> if (has_aux(event) && has_aux(output_event) &&
>>> - event->pmu != output_event->pmu)
>>> + event->pmu->setup_aux != output_event->pmu->setup_aux)
>>
>> It is not very flexible and risks someone adding aux PMUs that
>> do not want that rule but accidentally support it. Another
>> option is to add a PMU callback, but really you need to Peter's
>> feedback.
>
> Thanks a lot for sharing opinion, Adrian!
>
> How about below code? An alternative way is to compare the PMU's parent
> device driver, e.g. for Arm SPE PMU events, this can compare if two PMU
> events are using the Arm SPE driver.

IMHO, in the general case, whether 2 AUX area events can
output to the same buffer isn't really related to the device
hierarchy, driver or module.

>
> /*
> * If both events generate aux data, they must be on the same PMU
> * module but can be with different PMU instances.
> */
> if (has_aux(event) && has_aux(output_event)) {
> /* It isn't allowed if it fails to find driver pointer */
> if (!event->pmu->parent || !event->pmu->parent->driver)
> goto out;
>
> if (!output_event->pmu->parent || !output_event->pmu->parent->driver)
> goto out;
>
> /*
> * It isn't allowed if aux events are not same type of PMU
> * device. This is determined by comparing the associated
> * driver pointers.
> */
> if (event->pmu->parent->driver != output_event->pmu->parent->driver)
> goto out;
> }
>
> I verified the code above, it also works well at my side.
>
> @Peter.Z, Please let me know if this is okay for you.
>
> Thanks,
> Leo