Re: [PATCH v4 5/6] perf vendor events arm64: Update stall_slot workaround for N2 r0p3

[James Clark]

On 09/08/2023 14:54, John Garry wrote:
> On 09/08/2023 14:06, James Clark wrote:
>>> "MetricExpr": "(op_retired / op_spec) * (1 - (stall_slot if
>>> (cpuid_less_than(410fd493)) else (stall_slot - cpu_cycles)) / (#slots *
>>> cpu_cycles))"
>>>
>>> I'm currently figuring out how cpuid_less_than() would be implemented
>>> (I'm no great python wrangler), but it would be along the lines of what
>>> Ian added for "has_event" in
>>> https://urldefense.com/v3/__https://lore.kernel.org/linux-perf-users/20230623151016.4193660-1-irogers@xxxxxxxxxx/__;!!ACWV5N9M2RV99hQ!PlOppEWtIj9jDW2Zlon0zRZVpzPTzPvm5Ho5NnRIN0vD78iFcEzMEAtsrW_MrRPiW84XhWpbhc3seQcmLu-BfQ$ ;
>>> Thanks,
>>> John
>> Yeah it looks like it could be done that way. Also, the way I added it,
>> it doesn't have access to the PMU type, it just does a generic
>> pmu__find_core_pmu() so won't work very well for heterogeneous systems.
> 
> I haven't been keeping a close eye on the hybrid PMU support, but AFAIK
> metrics for hybrid arm64 system, i.e. bL, aren't supported - maybe that
> has changed. The gating for bL support was in pmu__find_core_pmu()
> returning NULL for a hybrid system.
> 

Yes we're not currently publishing any metrics for hybrid systems, so in
this case for N2 and V2 it's not needed. But it would be good to try to
future proof it if possible. Although I don't know how the metrics
currently react on hybrid systems, it's also something I have to take a
look at.


>>
>> If we're going to do a deeper modification of the expression parser like
>> with has_event() it might be possible to pass in the actual CPU ID that
>> the metric is running on which would be better.
>>
>> I'll have a look.
> 
> Thanks. I was playing with this yesterday, but I was making slow
> progress. I was essentially following the has_event example, but the
> argument type causes an issue, in that has_event expected an event name,
> while we want to pass a hex string.
> 
> If you could check this then that would be great.
> 
> Thanks,
> John
>