Re: [PATCH 1/2] perf script: Fix perf script -F +metric

[Artem Savkov <asavkov@redhat.com>]

On Fri, Jun 28, 2024 at 12:19:59PM -0700, Andi Kleen wrote:
> > This works, however there seems to be yet another issue on
> > perf-tools-next branch that breaks script metrics.
> >
> > 617824a7f0f73 (perf parse-events: Prefer sysfs/JSON hardware
> > events over legacy, 2024-04-15) changes event priority so on
> > my host "instructions" have type PERF_TYPE_RAW instead of
> > PERF_TYPE_HARDWARE.
>
> FWIW I don't see this on my build on Intel, even with that commit.
>
> > This results in evsel__stat_type no longer recognizing these as
> > STAT_INSTRUCTIONS.
>
> I guess we could just fix evsel__stat_type to handle that case
>
> Does this patch work?

It does for this specific example, but there are other events as well.
From arch/x86/events/intel/core.c:

        static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
        {
                [PERF_COUNT_HW_CPU_CYCLES]		= 0x003c,
                [PERF_COUNT_HW_INSTRUCTIONS]		= 0x00c0,
                [PERF_COUNT_HW_CACHE_REFERENCES]	= 0x4f2e,
                [PERF_COUNT_HW_CACHE_MISSES]		= 0x412e,
                [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x00c4,
                [PERF_COUNT_HW_BRANCH_MISSES]		= 0x00c5,
                [PERF_COUNT_HW_BUS_CYCLES]		= 0x013c,
                [PERF_COUNT_HW_REF_CPU_CYCLES]		= 0x0300, /* pseudo-encoding */
        };

The list is probably different on other arches.

> perf: Workaround event priority problems
>
> When instructions is not of HARDWARE types anymore, but raw,
> which breaks the shadow stat metrics detection for IPC.
>
> Reported-by: Artem Savkov <asavkov@redhat.com>
> Fixes: 617824a7f0f73 (perf parse-events: Prefer sysfs/JSON hardware
> Signed-off-by: Andi Kleen <ak@linux.intel.com>
>
> diff --git a/tools/perf/util/stat-shadow.c b/tools/perf/util/stat-shadow.c
> index 3466aa952442..87a503e47ff0 100644
> --- a/tools/perf/util/stat-shadow.c
> +++ b/tools/perf/util/stat-shadow.c
> @@ -97,9 +97,11 @@ static enum stat_type evsel__stat_type(const struct evsel *evsel)
>
>  	if (evsel__is_clock(evsel))
>  		return STAT_NSECS;
> -	else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES))
> +	else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES) ||
> +			!strcmp(evsel__name((struct evsel *)evsel), "cycles"))

It is also probably worth to add a PERF_TYPE_RAW check.

>  		return STAT_CYCLES;
> -	else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS))
> +	else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS) ||
> +			!strcmp(evsel__name((struct evsel *)evsel), "instructions"))
>  		return STAT_INSTRUCTIONS;
>  	else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
>  		return STAT_STALLED_CYCLES_FRONT;
>

--
 Artem