Re: [PATCH v3 3/9] perf annotate: Fix wrong --show-total-period option showing number of samples

From: Arnaldo Carvalho de Melo
Date: Mon Jul 24 2017 - 13:35:36 EST


Em Sun, Jul 23, 2017 at 08:46:20AM -0700, Andi Kleen escreveu:
> On Sun, Jul 23, 2017 at 07:46:05AM +0900, Namhyung Kim wrote:
> > Hi Arnaldo and Taeung,
> >
> > (+ Andi)
> >
> > On Fri, Jul 21, 2017 at 11:47:48AM -0300, Arnaldo Carvalho de Melo wrote:
> > > Em Thu, Jul 20, 2017 at 06:36:55AM +0900, Taeung Song escreveu:
> > > > +++ b/tools/perf/builtin-annotate.c
> > > > @@ -177,14 +177,12 @@ static int perf_evsel__add_sample(struct perf_evsel *evsel,
> > > > */
> > > > process_branch_stack(sample->branch_stack, al, sample);
> > > >
> > > > - sample->period = 1;
> > > > sample->weight = 1;
> > > > -
> > > > he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
> > > > if (he == NULL)
> > > > return -ENOMEM;
> > >
> > > I split the hunk above into a separate patch, as a fix, Namhyung, can
> > > you take a look at why need to unconditionally overwrite what is in
> > > sample->weight as well?
> > >
> > > Looks fishy as it may come with a value from the kernel, parsed in
> > > perf_evsel__parse_sample(), when PERF_SAMPLE_WEIGHT is in
> > > perf_event_attr->sample_type.
> > >
> > > Is it that the hists code needs a sane value when PERF_SAMPLE_WEIGHT
> > > isn't requested in sample_type?
> >
> > It was Andi added that code originally (05484298cbfe). IIUC the
> > weight is only meaningful for some CPUs with Intel TSX and he used a
> > dummy value.
>
> It's used for more than TSX. e.g. perf mem uses it for memory latencies.
>
> > AFAIK the hists code doesn't care of it unless weight sort key is used
> > (for report). As it's not used by annotate code, I think it'd be
> > better leaving it as is (like period).
>
> Right, it's needed when weight is specified as a sort key. But we need
> a fallback in case the user specified weight in perf report, but
> didn't enable it for perf record.

Humm, shouldn't we fail in that case? I.e. user asks for per-sample
property not collected at 'perf record' time?

That or the weight sort order handler should see that
perf_sample->weight is zero and assume it wasn't collected then turn it
into a 1? Or just look at evsel->attr.sample_type & PERF_SAMPLE_WEIGHT?

- Arnaldo