Re: [PATCH V3 1/3] perf/x86: Add new event for AUX output counter index

[Adrian Hunter]

On 10/09/21 7:29 pm, Liang, Kan wrote:
> 
> 
> On 9/10/2021 12:04 PM, Peter Zijlstra wrote:
>> On Tue, Sep 07, 2021 at 01:45:22PM -0400, Liang, Kan wrote:
>>> On 9/7/2021 12:39 PM, Adrian Hunter wrote:
>>
>>>> @@ -4494,8 +4500,16 @@ static int intel_pmu_check_period(struct perf_event *event, u64 value)
>>>>        return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0;
>>>>    }
>>>> +static void intel_aux_output_init(void)
>>>> +{
>>>> +    /* Refer also intel_pmu_aux_output_match() */
>>>> +    if (x86_pmu.intel_cap.pebs_output_pt_available)
>>>> +        x86_pmu.assign = intel_pmu_assign_event;
>>>> +}
>>>
>>> For a hybrid machine, x86_pmu.intel_cap.pebs_output_pt_available is always
>>> cleared. We probably need the PMU specific
>>> pmu->intel_cap.pebs_output_pt_available here.
>>>
>>>> +
>>>>    static int intel_pmu_aux_output_match(struct perf_event *event)
>>>>    {
>>>> +    /* intel_pmu_assign_event() is needed, refer intel_aux_output_init() */
>>>>        if (!x86_pmu.intel_cap.pebs_output_pt_available)
>>>>            return 0;
>>>>
>>>
>>> For a hybrid machine, this always return 0. I think we need to fix it first?
>>
>> AFAICT the patch is correct for !hybrid, and the hybrid PT muck can then
>> also fix this up, right?
>>
> 
> Yes, for !hybrid, the patch is good.
> 
> Since PEBS via PT is temporarily disabled for hybrid for now, the patch set should not bring any issues with hybrid either.
> The hybrid PT can be fixed separately.

I don't have much time to look at the hybrid case right now.

Would it be OK to go ahead with these patches?