Date:   Wed, 3 Nov 2021 08:24:03 +0100
From:   Jiri Olsa <jolsa@...hat.com>
To:     Namhyung Kim <namhyung@...nel.org>
Cc:     Arnaldo Carvalho de Melo <acme@...nel.org>,
        Ingo Molnar <mingo@...nel.org>,
        Peter Zijlstra <peterz@...radead.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Andi Kleen <ak@...ux.intel.com>,
        Ian Rogers <irogers@...gle.com>,
        Stephane Eranian <eranian@...gle.com>
Subject: Re: [PATCH v3] perf evsel: Fix missing exclude_{host,guest} setting

On Tue, Nov 02, 2021 at 04:21:21PM -0700, Namhyung Kim wrote:
> Hi Jiri,
> 
> On Tue, Nov 2, 2021 at 7:10 AM Jiri Olsa <jolsa@...hat.com> wrote:
> >
> > On Fri, Oct 29, 2021 at 03:49:29PM -0700, Namhyung Kim wrote:
> > > The current logic for the perf missing feature has a bug that it can
> > > wrongly clear some modifiers like G or H.  Actually some PMUs don't
> > > support any filtering or exclusion while others do.  But we check it
> > > as a global feature.
> > >
> > > For example, the cycles event can have 'G' modifier to enable it only
> > > in the guest mode on x86.  When you don't run any VMs it'll return 0.
> > >
> > >   # perf stat -a -e cycles:G sleep 1
> > >
> > >     Performance counter stats for 'system wide':
> > >
> > >                     0      cycles:G
> > >
> > >           1.000721670 seconds time elapsed
> > >
> > > But when it's used with other pmu events that don't support G modifier,
> > > it'll be reset and return non-zero values.
> > >
> > >   # perf stat -a -e cycles:G,msr/tsc/ sleep 1
> > >
> > >     Performance counter stats for 'system wide':
> > >
> > >           538,029,960      cycles:G
> > >        16,924,010,738      msr/tsc/
> > >
> > >           1.001815327 seconds time elapsed
> > >
> > > This is because of the missing feature detection logic being global.
> > > Add a hashmap to set pmu-specific exclude_host/guest features.
> > >
> > > Reported-by: Stephane Eranian <eranian@...gle.com>
> > > Signed-off-by: Namhyung Kim <namhyung@...nel.org>
> > > ---
> > > v3 changes)
> > >  * check memory allocation failure
> > >  * add more NULL check
> >
> > we were discussing this with Arnaldo yesterday and he had an idea to use
> > evsel->pmu link to store this info instead of hash.. I first thought we
> > needed 'evsel' related data, but after I gave it some thought I think that
> > might actually work
> 
> I don't get it.. do we have evsel->pmu already?  Or do you want to add it?
> Yeah, the filtering facility (attr.exclude_*) should be kept in a PMU data
> not in the evsel.  So I added a hashmap to find the pmu data from attr.type.
> How do I use evsel->pmu to store the info then?

evsel->pmu is not there yet (only evsel->pmu_name) so that
would need to be added.. we have evsel__find_pmu available

then the idea is to use evsel->pmu instead of the hasmap,
like add:

  struct pmu {
    ...
    bool missing_exclude_guest;
  };

set it when the guest filtering fails and and check it
instead of the hashmap__find call

> 
> >
> > my argument was following usecase:
> >
> >   cycles:G,instructions:G,pmu/bla1/:G,pmu/bla2/
> >
> > that we would falsely clear pmu/bla1/:G if we used the 'evsel->pmu' data..
> > but then I realized it's detection if pmu support :G and so if the :G is
> > not there, none of the events should have it
> >
> > thoughts?
> 
> I don't think I'm following well... ;-p
> 
> If the pmu doesn't support host/guest filtering, pmu/bla1/G
> may count something.  Not sure if it's better to error out.
> But the cycles:G and instructions:G should result in 0
> in case there's no VM running.

hm, I think if pmu doesn't support host/guest filtering then
I think 'pmu/bla1/G' should error, no? better no number than
bad number

jitka

> 
> Thanks,
> Namhyung
> 

Powered by blists - more mailing lists