[PATCH v3 2/7] perf metric: Event "Compat" value supports matching multiple identifiers

Jing Zhang renyu.zj at linux.alibaba.com
Mon Jun 19 19:12:58 PDT 2023



在 2023/6/19 下午5:31, John Garry 写道:
> On 19/06/2023 09:59, Jing Zhang wrote:
>>> Please verify running these metrics with 'perf stat', like 'perf stat -v -M slc_miss_rate'
>>>
>> Ok, it shows:
>> #./perf stat -v -M slc_miss_rate sleep 1
>>
>> metric expr hnf_cache_miss / hnf_slc_sf_cache_access for slc_miss_rate
>> found event duration_time
>> found event hnf_slc_sf_cache_access
> 
> In the earlier RFC series you had tools/perf/pmu-events/arch/arm64/arm/cmn700/sys/cmn.json, which describes event hnf_slc_sf_cache_access
> 
> But that JSON is not in this series. Why is it not included?
> 

Because the RFC version of the cmn.json file does not define the EventCode for each event, this will not take effect, so I temporarily removed it.
The EventID of CMN events is different from other events. For example, hnf_slc_sf_cache_access corresponds to arm_cmn_0/type=0x5,eventid=0x2/.
The current JSON format parsing does not support this EventID, and jevent.py needs to be extended.

> The cmn kernel driver exposes event hnf_slc_sf_cache_access, but I did not think that perf tool metric code matches those events described in /bus/event_sourcs/devices/<PMU>/events
> 

If there is no alias defined, other events with the same name may be matched, it is indeed necessary to define an alias for each event first,
and I will add it in the next version. But first I need to extend jevent.py to support cmn EventID.

>> found event hnf_cache_miss
>> Parsing metric events '{hnf_slc_sf_cache_access/metric-id=hnf_slc_sf_cache_access/,hnf_cache_miss/metric-id=hnf_cache_miss/}:W,duration_time'
>> hnf_slc_sf_cache_access -> arm_cmn_0/type=0x5,eventid=0x2/
>> hnf_slc_sf_cache_access -> arm_cmn_1/type=0x5,eventid=0x2/
>> hnf_cache_miss -> arm_cmn_0/type=0x5,eventid=0x1/
>> hnf_cache_miss -> arm_cmn_1/type=0x5,eventid=0x1/
>> Control descriptor is not initialized
>> hnf_slc_sf_cache_access: 127615 1001344900 1001344900
>> hnf_cache_miss: 36829 1001344900 1001344900
>> hnf_slc_sf_cache_access: 131526 1001343540 1001343540
>> hnf_cache_miss: 40587 1001343540 1001343540
>> duration_time: 1001381687 1001381687 1001381687
>>
>>   Performance counter stats for 'system wide':
>>
>>             259,141      hnf_slc_sf_cache_access   #     29.9 %  slc_miss_rate
>>              77,416      hnf_cache_miss
>>       1,001,381,687 ns   duration_time
>>
>>         1.001381687 seconds time elapsed
>>
>>
>>
>> #./perf list
>> ...
>>   arm_cmn_0/hnf_cache_miss/                          [Kernel PMU event]
>>   arm_cmn_0/hnf_slc_sf_cache_access/                 [Kernel PMU event]
>> ...
>>   arm_cmn_1/hnf_cache_miss/                          [Kernel PMU event]
>>   arm_cmn_1/hnf_slc_sf_cache_access/                 [Kernel PMU event]
>> ...
>>
>>>> In the past, I always thought that the function of the alias was to explain the meaning of these events in the perf list.
>>>> Or maybe I'm missing something?
>>> Event aliases do give the ability to describe the event in perf list. But we can also run them for 'perf stat', like:
>>>
>>> ./perf list uncore
>>> List of pre-defined events (to be used in -e or -M):
>>>
>>>    uncore_cbox_0/clockticks/                          [Kernel PMU event]
>>>    uncore_cbox_1/clockticks/                          [Kernel PMU event]
>>>    uncore_imc/data_reads/                             [Kernel PMU event]
>>>    uncore_imc/data_writes/                            [Kernel PMU event]
>>>    uncore_imc/gt_requests/                            [Kernel PMU event]
>>>    uncore_imc/ia_requests/                            [Kernel PMU event]
>>>    uncore_imc/io_requests/                            [Kernel PMU event]
>>>
>>> uncore cache:
>>>    unc_cbo_cache_lookup.any_es
>>>         [L3 Lookup any request that access cache and found line in E or S-state. Unit: uncore_cbox]
>>> ...
>>>
>>> sudo ./perf stat -v -e unc_cbo_cache_lookup.any_es
>>> Using CPUID GenuineIntel-6-3D-4
>>> unc_cbo_cache_lookup.any_es -> uncore_cbox_0/event=0x34,umask=0x86/
>>> unc_cbo_cache_lookup.any_es -> uncore_cbox_1/event=0x34,umask=0x86/
>>> Control descriptor is not initialized
>>> ^Cunc_cbo_cache_lookup.any_es: 14361103 1853372468 1853372468
>>> unc_cbo_cache_lookup.any_es: 14322188 1853360415 1853360415
>>>
>>>   Performance counter stats for 'system wide':
>>>
>>>          14,361,103      unc_cbo_cache_lookup.any_es
>>>          14,322,188      unc_cbo_cache_lookup.any_es
>>>
>>>         1.853388227 seconds time elapsed
>>>
>> Ok, thanks. If I use events without a prefix, such as perf stat -e clockticks sleep 1, will this also work?
> 
> In this case, yes - it would work for uncore_cbox_0/clockticks/ and uncore_cbox_1/clockticks/
> 
> But you need to be careful to here - if another PMU has same event name, then it might also match.
> 

Ok, I got it.

Thanks,
Jing

> Thanks,
> John



More information about the linux-arm-kernel mailing list