Re: [PATCH] perf tools: Add SPE total latency as PERF_SAMPLE_WEIGHT

[Namhyung Kim]

Hi Leo,

On Tue, Nov 30, 2021 at 9:08 PM Leo Yan <leo.yan@xxxxxxxxxx> wrote:
>
> Hi Namhyung,
>
> On Tue, Nov 30, 2021 at 04:39:08PM -0800, Namhyung Kim wrote:
> > Use total latency info in the SPE counter packet as sample weight so
> > that we can see it in local_weight and (global) weight sort keys.
> >
> > Maybe we can use PERF_SAMPLE_WEIGHT_STRUCT to support ins_lat as well
> > but I'm not sure which latency it matches.  So just adding total
> > latency first.
> >
> > Cc: German Gomez <german.gomez@xxxxxxx>
> > Cc: Leo Yan <leo.yan@xxxxxxxxxx>
> > Cc: Mark Rutland <mark.rutland@xxxxxxx>
> > Signed-off-by: Namhyung Kim <namhyung@xxxxxxxxxx>
> > ---
> >  tools/perf/util/arm-spe-decoder/arm-spe-decoder.c | 2 ++
> >  tools/perf/util/arm-spe-decoder/arm-spe-decoder.h | 1 +
> >  tools/perf/util/arm-spe.c                         | 4 +++-
> >  3 files changed, 6 insertions(+), 1 deletion(-)
> >
> > diff --git a/tools/perf/util/arm-spe-decoder/arm-spe-decoder.c b/tools/perf/util/arm-spe-decoder/arm-spe-decoder.c
> > index 3fc528c9270c..5e390a1a79ab 100644
> > --- a/tools/perf/util/arm-spe-decoder/arm-spe-decoder.c
> > +++ b/tools/perf/util/arm-spe-decoder/arm-spe-decoder.c
> > @@ -179,6 +179,8 @@ static int arm_spe_read_record(struct arm_spe_decoder *decoder)
> >                               decoder->record.phys_addr = ip;
> >                       break;
> >               case ARM_SPE_COUNTER:
> > +                     if (idx == SPE_CNT_PKT_HDR_INDEX_TOTAL_LAT)
> > +                             decoder->record.latency = payload;
> >                       break;
> >               case ARM_SPE_CONTEXT:
> >                       decoder->record.context_id = payload;
> > diff --git a/tools/perf/util/arm-spe-decoder/arm-spe-decoder.h b/tools/perf/util/arm-spe-decoder/arm-spe-decoder.h
> > index 46a8556a9e95..69b31084d6be 100644
> > --- a/tools/perf/util/arm-spe-decoder/arm-spe-decoder.h
> > +++ b/tools/perf/util/arm-spe-decoder/arm-spe-decoder.h
> > @@ -33,6 +33,7 @@ struct arm_spe_record {
> >       enum arm_spe_sample_type type;
> >       int err;
> >       u32 op;
> > +     u32 latency;
> >       u64 from_ip;
> >       u64 to_ip;
> >       u64 timestamp;
> > diff --git a/tools/perf/util/arm-spe.c b/tools/perf/util/arm-spe.c
> > index 4748bcfe61de..a756325c72a7 100644
> > --- a/tools/perf/util/arm-spe.c
> > +++ b/tools/perf/util/arm-spe.c
> > @@ -317,6 +317,7 @@ static int arm_spe__synth_mem_sample(struct arm_spe_queue *speq,
> >       sample.addr = record->virt_addr;
> >       sample.phys_addr = record->phys_addr;
> >       sample.data_src = data_src;
> > +     sample.weight = record->latency;
>
> The latency can be used for branch operations as well, it's good to
> assign latency for branch samples in the function
> arm_spe__synth_branch_sample().

Yep, I'll update.

>
> With adding latency for branch sample, the change would be good for me:
>
> Reviewed-by: Leo Yan <leo.yan@xxxxxxxxxx>

Thanks for your review!
Namhyung


>
> >       return arm_spe_deliver_synth_event(spe, speq, event, &sample);
> >  }
> > @@ -980,7 +981,8 @@ arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
> >       attr.type = PERF_TYPE_HARDWARE;
> >       attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK;
> >       attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
> > -                         PERF_SAMPLE_PERIOD | PERF_SAMPLE_DATA_SRC;
> > +                         PERF_SAMPLE_PERIOD | PERF_SAMPLE_DATA_SRC |
> > +                         PERF_SAMPLE_WEIGHT;
> >       if (spe->timeless_decoding)
> >               attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
> >       else
> > --
> > 2.34.0.rc2.393.gf8c9666880-goog
> >