Re: perf-tools-next: Request for testing hybrid changes

[Ravi Bangoria <ravi.bangoria@amd.com>]

>> For heterogeneous ARM PMUs I believe this is exactly the path taken to
>> cause the perf_event_open to fail. The resolution is to add the bit to
>> the PMU's capabilities. I believe doing it here will suffice:
>> https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/tree/drivers/perf/arm_pmu.c?h=perf/core#n877
>> So:
>> ```
>> .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS | PERF_PMU_CAP_EXTENDED_REGS,
>> ```
>> becomes:
>> ```
>> .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS |
>> PERF_PMU_CAP_EXTENDED_REGS | PERF_PMU_CAP_EXTENDED_HW_TYPE,
>> ```
>> Now strictly speaking you only need this in the case of heterogeneous
>> PMUs, but that's the case for other bits here and I don't see special
>> case logic...
>>
> 
> Hi Ian,
> 
> I tested the above and it seems to work fine on Arm so I will send it
> shortly. Some of the legacy perf stat use cases are working better with
> it so it makes sense to me. I also don't see a reason to not always do
> it even for homogeneous PMUs. To be honest I don't see a reason not to
> remove it as a capability and have it as the default behaviour
> everywhere? It all happens in the core code so I'm not sure adding it
> per PMU device is necessary, but maybe there is a reason?

EventSelect on AMD is 12 bits where upper nibble is passed via config[35:32].
i.e. it conflicts with the semantics of PERF_PMU_CAP_EXTENDED_HW_TYPE. So,
please don't make it default :)

Thanks,
Ravi