Re: [PATCH RFC 1/1] perf stat: do not fatal if the leader is errored

From: Dongli Zhang
Date: Thu Sep 22 2022 - 14:10:46 EST




On 9/22/22 1:07 AM, Like Xu wrote:
> On 22/9/2022 3:10 pm, Dongli Zhang wrote:
>> There are three options to fix the issue.
>>
>> 1. Do not expose /sys/bus/event_source/devices/cpu/events/slots to
>> userspace so that pmu_have_event(pmu->name, "slots") returns false.
>
> IMO, the guest PMU driver should be fixed
> since it misrepresents emulated hardware capabilities in terms of slots.
>
>>
>> 2. Run cpuid at perf userspace and avoid using 'slots' if it is not
>> supported in cpuid.
>>
>> 3. Do not fatal perf if the leader is failed. Do not create events for an
>> evsel if its leader is already failed.
>
> We may also need this since it's easier and more agile to update the perf tool
> than the kernel code or KVM emulated capabilities.

Thank you very much for the suggestion. I may remove the RFC and re-send the
patch for perf userspace.

BTW, I used to try your prior patch at KVM side (suggested by Bert) but it does
not help in this case. The VM kernel is already aware that 'slots' is not
supported by the hypervisor.

https://lore.kernel.org/lkml/20220106032118.34459-1-likexu@xxxxxxxxxxx/T/

Dongli Zhang

>
>>
>> This RFC patch is with the 3rd option. Would you mind suggesting which
>> option is better?