Re: [PATCH v1] perf evlist: Avoid frequency mode for the dummy event

[Ian Rogers]

On Thu, Sep 21, 2023 at 10:36 PM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>
> On 21/09/23 22:26, Namhyung Kim wrote:
> > Hi,
> >
> > On Mon, Sep 18, 2023 at 11:00 PM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
> >>
> >> On 19/09/23 00:48, Ian Rogers wrote:
> >>> On Mon, Sep 18, 2023 at 1:14 AM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
> >>>>
> >>>> On 16/09/23 07:09, Ian Rogers 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@xxxxxxxxxx>
> >>>>> Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
> >
> > I'll take the original patch first.
> >
> >
> >>>>> ---
> >>>>>  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");
> >>>>
> >>>> Note that evsel__config() will put it back to freq if -F is used.
> >>>
> >>> Right, I was looking for a minimal fix in part for the sake of back
> >>> porting. For the -F we could do:
> >>>
> >>> ```
> >>> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> >>> index d5363d23f5d3..806185a39e17 100644
> >>> --- a/tools/perf/util/evsel.c
> >>> +++ b/tools/perf/util/evsel.c
> >>> @@ -1083,11 +1083,15 @@ void __weak arch__post_evsel_config(struct
> >>> evsel *evsel __maybe_unused,
> >>> static void evsel__set_default_freq_period(struct record_opts *opts,
> >>>                                           struct perf_event_attr *attr)
> >>> {
> >>> -       if (opts->freq) {
> >>> +       bool is_dummy = attr->type == PERF_TYPE_SOFTWARE &&
> >>> +               attr->config == PERF_COUNT_SW_DUMMY;
> >>> +
> >>> +       if (opts->freq && !is_dummy) {
> >>>                attr->freq = 1;
> >>>                attr->sample_freq = opts->freq;
> >>>        } else {
> >>> -               attr->sample_period = opts->default_interval;
> >>> +               attr->freq = 0;
> >>> +               attr->sample_period = is_dummy ? 1 : opts->default_interval;
> >>>        }
> >>> }
> >>> ```
> >>>
> >>> But this felt like it could potentially have other side-effects.
> >>
> >> Perhaps leave it alone, if the period has already been defined:
> >>
> >> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> >> index d5363d23f5d3..ad3e12f5ec88 100644
> >> --- a/tools/perf/util/evsel.c
> >> +++ b/tools/perf/util/evsel.c
> >> @@ -1166,7 +1166,8 @@ void evsel__config(struct evsel *evsel, struct record_opts *opts,
> >>         if ((evsel->is_libpfm_event && !attr->sample_period) ||
> >>             (!evsel->is_libpfm_event && (!attr->sample_period ||
> >>                                          opts->user_freq != UINT_MAX ||
> >> -                                        opts->user_interval != ULLONG_MAX)))
> >> +                                        opts->user_interval != ULLONG_MAX) &&
> >> +            !(is_dummy && attr->sample_period)))
> >>                 evsel__set_default_freq_period(opts, attr);
> >>
> >>         /*
> >
> > Or simply like this?
> >
> >
> > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > index d5363d23f5d3..6ce832ce62f1 100644
> > --- a/tools/perf/util/evsel.c
> > +++ b/tools/perf/util/evsel.c
> > @@ -1169,6 +1169,9 @@ void evsel__config(struct evsel *evsel, struct
> > record_opts *opts,
> >                                          opts->user_interval != ULLONG_MAX)))
> >                 evsel__set_default_freq_period(opts, attr);
> >
> > +       if (evsel__is_dummy_event(evsel))
> > +               attr->freq = 0;
> > +
> >         /*
> >          * If attr->freq was set (here or earlier), ask for period
> >          * to be sampled.
>
> I thought there might be corner cases where it made a difference,
> but I can't find any, so that should do.
>

It seemed more intention revealing to do it at creation/initialization
than on a later not obviously executed code path - I'm thinking of
future me trying to understand the code. My priority is the clearing
of the flag, so I'm easy.

Thanks,
Ian