Re: [PATCH V2 1/6] perf/x86/intel: Add Grand Ridge and Sierra Forest

[Liang, Kan]

On 2023-06-06 3:37 p.m., Peter Zijlstra wrote:
> On Tue, Jun 06, 2023 at 02:34:47PM -0400, Liang, Kan wrote:
>>
>>
>> On 2023-06-06 2:17 p.m., Peter Zijlstra wrote:
>>> On Tue, Jun 06, 2023 at 12:16:29PM -0400, Liang, Kan wrote:
>>>
>>>>> names for a reason, so that enums like the above become something
>>>>> sensible like:
>>>>>
>>>>> 	case INTEL_FAM6_ATOM_CRESTMONT:
>>>>> 	case INTEL_FAM6_ATOM_CRESTMONT_X:
>>>>>
>>>>> and now it's super obvious why they're grouped.
>>>>>
>>>>>>> +		pr_cont("Crestmont events, ");
>>>>
>>>> The Sierra Forest should not be a platform name. I think it's the code
>>>> name of the processor.
>>>>
>>>> The problem is that the uarch name doesn't work for the hybrid, since it
>>>> has different uarchs in the same processors. To make the naming rules
>>>> consistent among big core, atom, and hybrid, maybe we should use the
>>>> code name of the processor in intel-family.h.
>>>
>>> I obviously disagree; these are not hybrid and calling them both
>>> CRESTMONT makes *far* more sense than the random gibberish they're
>>> called now.
>>>
>>> Yes, hybrid made a complete mess of things (in many ways), but we should
>>> then not do the obvious correct thing for when we can.
>>
>> Besides hybrid, it seems there is a bigger problem for the big core.
>>
>> The big core uses the processor code name since Ice Lake. In the perf
>> code, we also uses the processor code name for the big core.
>> 	pr_cont("Icelake events, ");
> 
> Yeah, it's a mess :/ Ideally that would print "Sunny Cove", but I think
> there's userspace looking at that string :-(

Yes, for the existing names, we probably cannot change it. I will try to
only use the micro-architecture name for the future platforms in perf.

> 
>> Is it OK to leave the big core as is (using processor code name), but
>> only change the name for Grand Ridge and Sierra Forest?
> 
> Arguably we should also rename ALDERLAKE_N to ATOM_GRACEMONT
> 
> 
> We should also do something about that whole hybrid init thing, the
> meteorlake stuff is quite a mess as well. Perhaps we can add hybrid_pmu
> next to intel_pmu to have a better start in life for x86_pmu.
>

I will think about it and try to clean up the hybrid init.

> Anyway, we should really try not to make a bigger mess and try and clean
> up where we can.

Sure.

Thanks,
Kan