[PATCH RFC 2/9] perf metrics: Don't iter sys metrics if we already found a CPU match

John Garry john.g.garry at oracle.com
Wed Jul 12 02:37:36 PDT 2023


On 12/07/2023 06:40, Ian Rogers wrote:
>> A few points to make on this:
>> - Currently we don't have any same-named metrics like this, so not much
>> use in supporting it in the code (yet).
> We have same named metrics for heterogeneous CPU PMUs:
> https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/arch/x86/alderlake/adl-metrics.json?h=perf-tools-next*n304__;Iw!!ACWV5N9M2RV99hQ!MyvM7oyC6FgOVgDn2-Ot_TJNh4TF_VM9SlIVwv2AOTkJGdmDJ2NYf5WXh-yLcG1dRxLKdXWZVTzsoOo5yDk$  
> cpu_atom
> https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/arch/x86/alderlake/adl-metrics.json?h=perf-tools-next*n1125__;Iw!!ACWV5N9M2RV99hQ!MyvM7oyC6FgOVgDn2-Ot_TJNh4TF_VM9SlIVwv2AOTkJGdmDJ2NYf5WXh-yLcG1dRxLKdXWZVTzsL1dVM3Q$  
> cpu_core
> 

I meant that we have no same-named events for sys PMUs compared to 
uncore/CPU PMUs.

>> - Even if we had some same-named metrics, I am not sure if it even works
>> properly. Do we have any uncore PMU metrics which have same name as CPU
>> metrics?
> So I was thinking IPC was a generic concept that would apply to a
> co-processor on a network card, a GPU, etc.
> 
>> - Further to the previous point, do we really want same-named metrics
>> for different PMUs in the future? I think event / metric names need to
>> be chosen carefully to avoid clash for other PMUs or keywords. For your
>> example, if I did ask for IPC metric, I'd like to be able to just know
>> I'm getting IPC metric for CPUs or some other PMUs, but not both.
> At the moment if you request an event without a PMU, say instructions
> retired, we will attempt to open the event on every PMU - legacy
> events (PERF_TYPE_HARDWARE, PERF_TYPE_HW_CACHE) only try the core
> PMUs. It would seem consistent if metrics tried to open on every PMU
> like most events.

OK, fine. I can drop this change if you prefer. But, to reiterate my 
main point, I still think that there is not much point in looking for 
metrics which currently would not exist.

Thanks,
John




More information about the linux-arm-kernel mailing list