RE: [RFC PATCH v3 13/18] perf stat: Handle taken alone in hardware-grouping

[Wang, Weilin]

> -----Original Message-----
> From: Ian Rogers <irogers@xxxxxxxxxx>
> Sent: Wednesday, January 24, 2024 9:03 PM
> To: Wang, Weilin <weilin.wang@xxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>; Ingo Molnar <mingo@xxxxxxxxxx>;
> Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>; Alexander Shishkin
> <alexander.shishkin@xxxxxxxxxxxxxxx>; Jiri Olsa <jolsa@xxxxxxxxxx>; Namhyung
> Kim <namhyung@xxxxxxxxxx>; Hunter, Adrian <adrian.hunter@xxxxxxxxx>;
> Kan Liang <kan.liang@xxxxxxxxxxxxxxx>; linux-perf-users@xxxxxxxxxxxxxxx;
> linux-kernel@xxxxxxxxxxxxxxx; Taylor, Perry <perry.taylor@xxxxxxxxx>; Alt,
> Samantha <samantha.alt@xxxxxxxxx>; Biggers, Caleb
> <caleb.biggers@xxxxxxxxx>; Mark Rutland <mark.rutland@xxxxxxx>; Yang
> Jihong <yangjihong1@xxxxxxxxxx>
> Subject: Re: [RFC PATCH v3 13/18] perf stat: Handle taken alone in hardware-
> grouping
> 
> On Tue, Dec 12, 2023 at 3:03 PM <weilin.wang@xxxxxxxxx> wrote:
> >
> > From: Weilin Wang <weilin.wang@xxxxxxxxx>
> >
> > Add taken alone into consideration when grouping. Only one taken
> > alone event is supported per group.
> 
> This seems to be a deficiency also in
> parse_events__sort_events_and_fix_groups:
> https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-
> next.git/tree/tools/perf/util/parse-events.c?h=perf-tools-next#n2025
> Is there an example of events that must be taken alone? Perhaps
> because they share a fixed counter?

Yes, there are a few events must be taken alone. I did a quick search and 
found 29 events in SPR that are taken alone. The event
INT_MISC.UNKNOWN_BRANCH_CYCLES is an example. It is used in 
tma_unknown_branches. Usually taken alone events do not use fixed 
counter. INT_MISC.UNKNOWN_BRANCH_CYCLES could use any of the GP 
counters on SPR.

Taken alone events could be grouped with other non-taken alone events. 
But two taken alone events cannot be in the same group. 

For example, UOPS_RETIRED.MS is another taken alone event in SPR. If we 
do "perf stat -e UOPS_RETIRED.MS -e INT_MISC.UNKNOWN_BRANCH_CYCLES",
we will see multiplexing of 50% on each event. If we run "perf stat
 -e INT_MISC.CLEAR_RESTEER_CYCLES -e INT_MISC.UNKNOWN_BRANCH_CYCLES",
where INT_MISC.CLEAR_RESTEER_CYCLES is not taken alone, then we don't
have any multiplexing. 


> 
> Thanks,
> Ian
> 
> > Signed-off-by: Weilin Wang <weilin.wang@xxxxxxxxx>
> > ---
> >  tools/perf/pmu-events/jevents.py   |  7 +++++--
> >  tools/perf/pmu-events/pmu-events.h |  1 +
> >  tools/perf/util/metricgroup.c      | 18 +++++++++++++-----
> >  tools/perf/util/metricgroup.h      |  2 ++
> >  4 files changed, 21 insertions(+), 7 deletions(-)
> >
> > diff --git a/tools/perf/pmu-events/jevents.py b/tools/perf/pmu-
> events/jevents.py
> > index 16f17d335f8e..e5919e5e6d4a 100755
> > --- a/tools/perf/pmu-events/jevents.py
> > +++ b/tools/perf/pmu-events/jevents.py
> > @@ -55,7 +55,9 @@ _json_event_attributes = [
> >      # Counter this event could use
> >      'counter',
> >      # Longer things (the last won't be iterated over during decompress).
> > -    'long_desc'
> > +    'long_desc',
> > +    # Taken alone event could not be collected in the same group with other
> taken alone event
> > +    'taken_alone'
> >  ]
> >
> >  # Attributes that are in pmu_unit_layout.
> > @@ -71,7 +73,7 @@ _json_metric_attributes = [
> >  ]
> >  # Attributes that are bools or enum int values, encoded as '0', '1',...
> >  _json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping',
> 'perpkg',
> > -    'size', 'fixed_size'
> > +    'size', 'fixed_size', 'taken_alone'
> >  ]
> >
> >  def removesuffix(s: str, suffix: str) -> str:
> > @@ -331,6 +333,7 @@ class JsonEvent:
> >        extra_desc += '  Spec update: ' + jd['Errata']
> >      self.pmu = unit_to_pmu(jd.get('Unit'))
> >      self.counter = jd.get('Counter')
> > +    self.taken_alone = jd.get('TakenAlone')
> >      self.size = jd.get('Size')
> >      self.fixed_size = jd.get('FixedSize')
> >      filter = jd.get('Filter')
> > diff --git a/tools/perf/pmu-events/pmu-events.h b/tools/perf/pmu-
> events/pmu-events.h
> > index 65e0c5dd8bb4..ce10356e9665 100644
> > --- a/tools/perf/pmu-events/pmu-events.h
> > +++ b/tools/perf/pmu-events/pmu-events.h
> > @@ -50,6 +50,7 @@ struct pmu_event {
> >         const char *unit;
> >         bool perpkg;
> >         bool deprecated;
> > +       bool taken_alone;
> >  };
> >
> >  struct pmu_metric {
> > diff --git a/tools/perf/util/metricgroup.c b/tools/perf/util/metricgroup.c
> > index 016ea949ba5d..78e607f3ad56 100644
> > --- a/tools/perf/util/metricgroup.c
> > +++ b/tools/perf/util/metricgroup.c
> > @@ -1624,6 +1624,7 @@ static void metricgroup__free_pmu_info(struct
> list_head *pmu_info_list)
> >  static struct metricgroup__event_info *event_info__new(const char *name,
> >                                                       const char *pmu_name,
> >                                                       const char *counter,
> > +                                                     bool taken_alone,
> >                                                       bool free_counter)
> >  {
> >         int ret = 0;
> > @@ -1638,6 +1639,7 @@ static struct metricgroup__event_info
> *event_info__new(const char *name,
> >                 pmu_name = "core";
> >
> >         e->name = name;
> > +       e->taken_alone = taken_alone;
> >         e->free_counter = free_counter;
> >         e->pmu_name = pmu_name;
> >         if (free_counter) {
> > @@ -1676,7 +1678,8 @@ static int
> metricgroup__add_metric_event_callback(const struct pmu_event *pe,
> >         if (!strcasecmp(pe->name, d->event_name)) {
> >                 if (!pe->counter)
> >                         return -EINVAL;
> > -               event = event_info__new(d->event_id, pe->pmu, pe->counter,
> /*free_counter=*/false);
> > +               event = event_info__new(d->event_id, pe->pmu, pe->counter,
> > +                                       pe->taken_alone, /*free_counter=*/false);
> >                 if (!event)
> >                         return -ENOMEM;
> >                 list_add(&event->nd, d->list);
> > @@ -1799,6 +1802,8 @@ static int find_and_set_counters(struct
> metricgroup__event_info *e,
> >         int ret;
> >         unsigned long find_bit = 0;
> >
> > +       if (e->taken_alone && current_group->taken_alone)
> > +               return -ENOSPC;
> >         if (e->free_counter)
> >                 return 0;
> >         if (e->fixed_counter) {
> > @@ -1833,11 +1838,13 @@ static int _insert_event(struct
> metricgroup__event_info *e,
> >                 list_add(&event->nd, &group->event_head);
> >         else
> >                 list_add_tail(&event->nd, &group->event_head);
> > +       if (e->taken_alone)
> > +               group->taken_alone = true;
> >         return 0;
> >  }
> >
> >  /**
> > - * Insert the new_group node at the end of the group list.
> > + * Initialize the new group and insert it to the end of the group list.
> >   */
> >  static int insert_new_group(struct list_head *head,
> >                            struct metricgroup__group *new_group,
> > @@ -1847,6 +1854,7 @@ static int insert_new_group(struct list_head
> *head,
> >         INIT_LIST_HEAD(&new_group->event_head);
> >         fill_counter_bitmap(new_group->gp_counters, 0, size);
> >         fill_counter_bitmap(new_group->fixed_counters, 0, fixed_size);
> > +       new_group->taken_alone = false;
> >         list_add_tail(&new_group->nd, head);
> >         return 0;
> >  }
> > @@ -2050,8 +2058,8 @@ static int create_grouping(struct list_head
> *pmu_info_list,
> >         //TODO: for each new core group, we should consider to add events that
> uses fixed counters
> >         list_for_each_entry(e, event_info_list, nd) {
> >                 bitmap_scnprintf(e->counters, NR_COUNTERS, bit_buf,
> NR_COUNTERS);
> > -               pr_debug("Event name %s, [pmu]=%s, [counters]=%s\n", e->name,
> > -                       e->pmu_name, bit_buf);
> > +               pr_debug("Event name %s, [pmu]=%s, [counters]=%s,
> [taken_alone]=%d\n",
> > +                       e->name, e->pmu_name, bit_buf, e->taken_alone);
> >                 ret = assign_event_grouping(e, pmu_info_list, &groups);
> >                 if (ret)
> >                         goto out;
> > @@ -2098,7 +2106,7 @@ static int hw_aware_build_grouping(struct
> expr_parse_ctx *ctx __maybe_unused,
> >                 if (is_special_event(id)) {
> >                         struct metricgroup__event_info *event;
> >
> > -                       event = event_info__new(id, "default_core", "0",
> > +                       event = event_info__new(id, "default_core", "0", false,
> >                                                 /*free_counter=*/true);
> >                         if (!event)
> >                                 goto err_out;
> > diff --git a/tools/perf/util/metricgroup.h b/tools/perf/util/metricgroup.h
> > index 37fdf8c5ba77..8f6a1bb9d1ad 100644
> > --- a/tools/perf/util/metricgroup.h
> > +++ b/tools/perf/util/metricgroup.h
> > @@ -105,6 +105,7 @@ struct metricgroup__event_info {
> >         const char *pmu_name;
> >         bool fixed_counter;
> >         bool free_counter;
> > +       bool taken_alone;
> >         /** The counters the event allowed to be collected on. */
> >         DECLARE_BITMAP(counters, NR_COUNTERS);
> >  };
> > @@ -154,6 +155,7 @@ struct metricgroup__group {
> >         DECLARE_BITMAP(fixed_counters, NR_COUNTERS);
> >         /** Head to the list of event names in this group*/
> >         struct list_head event_head;
> > +       bool taken_alone;
> >  };
> >
> >  struct metricgroup__group_events {
> > --
> > 2.39.3
> >