Date:   Wed, 20 Sep 2023 22:04:33 -0700
From:   Mingwei Zhang <mizhang@...gle.com>
To:     Ian Rogers <irogers@...gle.com>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...hat.com>,
        Arnaldo Carvalho de Melo <acme@...nel.org>,
        Mark Rutland <mark.rutland@....com>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Jiri Olsa <jolsa@...nel.org>,
        Namhyung Kim <namhyung@...nel.org>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Kan Liang <kan.liang@...ux.intel.com>,
        linux-perf-users@...r.kernel.org, linux-kernel@...r.kernel.org,
        Yang Jihong <yangjihong1@...wei.com>,
        Stephane Eranian <eranian@...gle.com>
Subject: Re: [PATCH v1] perf evlist: Avoid frequency mode for the dummy event

On Mon, Sep 18, 2023 at 3:43 PM Ian Rogers <irogers@...gle.com> wrote:
>
> On Sat, Sep 16, 2023 at 5:46 PM Mingwei Zhang <mizhang@...gle.com> wrote:
> >
> > On Fri, Sep 15, 2023 at 9:10 PM Ian Rogers <irogers@...gle.com> wrote:
> > >
> > > Dummy events are created with an attribute where the period and freq
> > > are zero. evsel__config will then see the uninitialized values and
> > > initialize them in evsel__default_freq_period. As fequency mode is
> > > used by default the dummy event would be set to use frequency
> > > mode. However, this has no effect on the dummy event but does cause
> > > unnecessary timers/interrupts. Avoid this overhead by setting the
> > > period to 1 for dummy events.
> > >
> > > evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> > > period=1. This isn't necessary after this change and so the setting is
> > > removed.
> > >
> > > From Stephane:
> > >
> > > The dummy event is not counting anything. It is used to collect mmap
> > > records and avoid a race condition during the synthesize mmap phase of
> > > perf record. As such, it should not cause any overhead during active
> > > profiling. Yet, it did. Because of a bug the dummy event was
> > > programmed as a sampling event in frequency mode. Events in that mode
> > > incur more kernel overheads because on timer tick, the kernel has to
> > > look at the number of samples for each event and potentially adjust
> > > the sampling period to achieve the desired frequency. The dummy event
> > > was therefore adding a frequency event to task and ctx contexts we may
> > > otherwise not have any, e.g., perf record -a -e
> > > cpu/event=0x3c,period=10000000/. On each timer tick the
> > > perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> > > non-zero, then the kernel will loop over ALL the events of the context
> > > looking for frequency mode ones. In doing, so it locks the context,
> > > and enable/disable the PMU of each hw event. If all the events of the
> > > context are in period mode, the kernel will have to traverse the list for
> > > nothing incurring overhead. The overhead is multiplied by a very large
> > > factor when this happens in a guest kernel. There is no need for the
> > > dummy event to be in frequency mode, it does not count anything and
> > > therefore should not cause extra overhead for no reason.
> > >
> > > Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> > > Reported-by: Stephane Eranian <eranian@...gle.com>
> > > Signed-off-by: Ian Rogers <irogers@...gle.com>
> > > ---
> > >  tools/perf/util/evlist.c | 5 +++--
> > >  1 file changed, 3 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> > > index 25c3ebe2c2f5..e36da58522ef 100644
> > > --- a/tools/perf/util/evlist.c
> > > +++ b/tools/perf/util/evlist.c
> > > @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> > >                 .type   = PERF_TYPE_SOFTWARE,
> > >                 .config = PERF_COUNT_SW_DUMMY,
> > >                 .size   = sizeof(attr), /* to capture ABI version */
> > > +               /* Avoid frequency mode for dummy events to avoid associated timers. */
> > > +               .freq = 0,
> > > +               .sample_period = 1,
> > >         };
> > >
> > >         return evsel__new_idx(&attr, evlist->core.nr_entries);
> > > @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> > >         evsel->core.attr.exclude_kernel = 1;
> > >         evsel->core.attr.exclude_guest = 1;
> > >         evsel->core.attr.exclude_hv = 1;
> > > -       evsel->core.attr.freq = 0;
> > > -       evsel->core.attr.sample_period = 1;
> > >         evsel->core.system_wide = system_wide;
> > >         evsel->no_aux_samples = true;
> > >         evsel->name = strdup("dummy:u");
> > > --
> > > 2.42.0.459.ge4e396fd5e-goog
> > >
> >
> > Thank you very much for the change. I have one quick question about
> > the PMU unthrottling logic. When I am looking into the function
> > perf_adjust_freq_unthr_context(), I see the loop with PMU stop and
> > start in each iteration. Is there a good way to avoid this PMU reset
> > operation while quickly figuring out the event in frequency mode?
>
> Agreed. I think before the pmu_disable could be avoided for this condition:
> ```
> if (event->hw.interrupts != MAX_INTERRUPTS &&
>     (!event->attr.freq || !event->attr.sample_freq))
>         continue;
> ```
> Fixing up the event stop/start looks harder.
>

Right, I think putting the check early before pmu_disable() is already
a great optimization. The only concern I initially had was whether
event->hw.interrupts can be accessed before we disable the pmu. But
after checking this field in other locations, I don't see any problem
at all.

Thanks.
-Mingwei

Powered by blists - more mailing lists