Date:   Mon, 15 Nov 2021 17:01:40 -0800
From:   Ian Rogers <irogers@...gle.com>
To:     Jiri Olsa <jolsa@...hat.com>, Andi Kleen <ak@...ux.intel.com>,
        eranian@...gle.com, Kan Liang <kan.liang@...ux.intel.com>
Cc:     John Garry <john.garry@...wei.com>,
        Kajol Jain <kjain@...ux.ibm.com>,
        "Paul A . Clarke" <pc@...ibm.com>,
        Arnaldo Carvalho de Melo <acme@...nel.org>,
        Riccardo Mancini <rickyman7@...il.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...hat.com>,
        Mark Rutland <mark.rutland@....com>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        linux-perf-users@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 2/2] perf parse-events: Architecture specific leader override

On Sun, Nov 14, 2021 at 10:17 AM Ian Rogers <irogers@...gle.com> wrote:
>
> On Sun, Nov 14, 2021 at 9:02 AM Jiri Olsa <jolsa@...hat.com> wrote:
> >
> > On Fri, Nov 12, 2021 at 11:15:48PM -0800, Ian Rogers wrote:
> > > Currently topdown events must appear after a slots event:
> > >
> > > $ perf stat -e '{slots,topdown-fe-bound}' /bin/true
> > >
> > >  Performance counter stats for '/bin/true':
> > >
> > >          3,183,090      slots
> > >            986,133      topdown-fe-bound
> > >
> > > Reversing the events yields:
> > >
> > > $ perf stat -e '{topdown-fe-bound,slots}' /bin/true
> > > Error:
> > > The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (topdown-fe-bound).
> >
> > why is this actually failing?
> >
> > >
> > > For metrics the order of events is determined by iterating over a
> > > hashmap, and so slots isn't guaranteed to be first which can yield this
> > > error.
> > >
> > > Change the set_leader in parse-events, called when a group is closed, so
> > > that rather than always making the first event the leader, if the slots
> > > event exists then it is made the leader. It is then moved to the head of
> > > the evlist otherwise it won't be opened in the correct order.
> > >
> > > The result is:
> > >
> > > $ perf stat -e '{topdown-fe-bound,slots}' /bin/true
> > >
> > >  Performance counter stats for '/bin/true':
> > >
> > >          3,274,795      slots
> > >          1,001,702      topdown-fe-bound
> > >
> > > A problem with this approach is the slots event is identified by name,
> > > names can be overwritten like 'cpu/slots,name=foo/' and this causes the
> > > leader change to fail.
> >
> > would it be better then to move this shuffle to the metric code directly,
> > and make sure it only spits 'good' order of events?
> >
> > and keep "-e '{topdown-fe-bound,slots}'" to fail if user specifies that on
> > the command line
>
> It's an alternative to do that, but the people I spoke to preferred
> having parse-events do this. I'm not sure what Andi's opinion is.
> There is a general frustration about how brittle the slots event is,
> and so this does make it less brittle.

A different problem this fixes is if you have a group like
'{instructions,slots,topdown-fe-bound}' then slots still has to be the
group leader even though it is appearing before the topdown event. So
before the patch:

$ perf stat -e '{instructions,slots,topdown-fe-bound}' -a -- sleep 2

 Performance counter stats for 'system wide':

     <not counted>      instructions
     <not counted>      slots
   <not supported>      topdown-fe-bound

       2.006410898 seconds time elapsed

After:

$ perf stat -e '{instructions,slots,topdown-fe-bound}' -a -- sleep 2

 Performance counter stats for 'system wide':

        2572632080      slots
         362537420      instructions
         742416906      topdown-fe-bound

       2.005875050 seconds time elapsed

Thanks,
Ian

> > >
> > > Signed-off-by: Ian Rogers <irogers@...gle.com>
> > > ---
> > >  tools/perf/arch/x86/util/evlist.c | 17 +++++++++++++++++
> > >  tools/perf/util/evlist.h          |  1 +
> > >  tools/perf/util/parse-events.c    | 16 +++++++++++-----
> > >  tools/perf/util/parse-events.h    |  4 ++--
> > >  tools/perf/util/parse-events.y    | 12 ++++++++----
> > >  5 files changed, 39 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/tools/perf/arch/x86/util/evlist.c b/tools/perf/arch/x86/util/evlist.c
> > > index 0b0951030a2f..1624372b2da2 100644
> > > --- a/tools/perf/arch/x86/util/evlist.c
> > > +++ b/tools/perf/arch/x86/util/evlist.c
> > > @@ -17,3 +17,20 @@ int arch_evlist__add_default_attrs(struct evlist *evlist)
> > >       else
> > >               return parse_events(evlist, TOPDOWN_L1_EVENTS, NULL);
> > >  }
> > > +
> > > +struct evsel *arch_evlist__leader(struct list_head *list)
> > > +{
> > > +     struct evsel *evsel, *first;
> > > +
> > > +     first = list_entry(list->next, struct evsel, core.node);
> > > +
> > > +     if (!pmu_have_event("cpu", "slots"))
> > > +             return first;
> > > +
> > > +     __evlist__for_each_entry(list, evsel) {
> > > +             if (evsel->pmu_name && !strcmp(evsel->pmu_name, "cpu") &&
> > > +                     evsel->name && strstr(evsel->name, "slots"))
> > > +                     return evsel;
> > > +     }
> > > +     return first;
> > > +}
> > > diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
> > > index 97bfb8d0be4f..993437ffe429 100644
> > > --- a/tools/perf/util/evlist.h
> > > +++ b/tools/perf/util/evlist.h
> > > @@ -110,6 +110,7 @@ int __evlist__add_default_attrs(struct evlist *evlist,
> > >       __evlist__add_default_attrs(evlist, array, ARRAY_SIZE(array))
> > >
> > >  int arch_evlist__add_default_attrs(struct evlist *evlist);
> > > +struct evsel *arch_evlist__leader(struct list_head *list);
> > >
> > >  int evlist__add_dummy(struct evlist *evlist);
> > >
> > > diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
> > > index 6308ba739d19..a2f4c086986f 100644
> > > --- a/tools/perf/util/parse-events.c
> > > +++ b/tools/perf/util/parse-events.c
> > > @@ -1821,22 +1821,28 @@ parse_events__set_leader_for_uncore_aliase(char *name, struct list_head *list,
> > >       return ret;
> > >  }
> > >
> > > -void parse_events__set_leader(char *name, struct list_head *list,
> > > -                           struct parse_events_state *parse_state)
> > > +__weak struct evsel *arch_evlist__leader(struct list_head *list)
> > > +{
> > > +     return list_entry(list->next, struct evsel, core.node);
> > > +}
> > > +
> > > +struct list_head *parse_events__set_leader(char *name, struct list_head *list,
> > > +                                     struct parse_events_state *parse_state)
> > >  {
> > >       struct evsel *leader;
> > >
> > >       if (list_empty(list)) {
> > >               WARN_ONCE(true, "WARNING: failed to set leader: empty list");
> > > -             return;
> > > +             return NULL;
> > >       }
> > >
> > >       if (parse_events__set_leader_for_uncore_aliase(name, list, parse_state))
> > > -             return;
> > > +             return NULL;
> > >
> > > -     leader = list_entry(list->next, struct evsel, core.node);
> > > +     leader = arch_evlist__leader(list);
> > >       __perf_evlist__set_leader(list, &leader->core);
> > >       leader->group_name = name ? strdup(name) : NULL;
> > > +     return &leader->core.node;
> > >  }
> > >
> > >  /* list_event is assumed to point to malloc'ed memory */
> > > diff --git a/tools/perf/util/parse-events.h b/tools/perf/util/parse-events.h
> > > index c7fc93f54577..8a6e6b4d5cbe 100644
> > > --- a/tools/perf/util/parse-events.h
> > > +++ b/tools/perf/util/parse-events.h
> > > @@ -211,8 +211,8 @@ int parse_events_copy_term_list(struct list_head *old,
> > >
> > >  enum perf_pmu_event_symbol_type
> > >  perf_pmu__parse_check(const char *name);
> > > -void parse_events__set_leader(char *name, struct list_head *list,
> > > -                           struct parse_events_state *parse_state);
> > > +struct list_head *parse_events__set_leader(char *name, struct list_head *list,
> > > +                                     struct parse_events_state *parse_state);
> > >  void parse_events_update_lists(struct list_head *list_event,
> > >                              struct list_head *list_all);
> > >  void parse_events_evlist_error(struct parse_events_state *parse_state,
> > > diff --git a/tools/perf/util/parse-events.y b/tools/perf/util/parse-events.y
> > > index 174158982fae..5358c400f938 100644
> > > --- a/tools/perf/util/parse-events.y
> > > +++ b/tools/perf/util/parse-events.y
> > > @@ -199,20 +199,24 @@ group_def
> > >  group_def:
> > >  PE_NAME '{' events '}'
> > >  {
> > > -     struct list_head *list = $3;
> > > +     struct list_head *new_head, *list = $3;
> > >
> > >       inc_group_count(list, _parse_state);
> > > -     parse_events__set_leader($1, list, _parse_state);
> > > +     new_head = parse_events__set_leader($1, list, _parse_state);
> > > +     if (new_head)
> > > +             list_move(new_head, list);
> >
> > why not put these last 2 lines to parse_events__set_leader as well?
>
> I can move that, but I'll try to motivate having it this way. The list
> logic in Linux I find troublesome. In parse-events.y we have a list
> head that is just a next, prev pointer and not full evsel in a list.
> So for {topdown-fe-bound,slots} in parse-events.y we have a list of:
>
> head -next-> evsel("topdown-fe-bound") -next-> evsel("slots")
>
> But in __perf_evlist__set_leader the list doesn't have the head:
>
> evsel("topdown-fe-bound") -next-> evsel("slots")
>
> The list_move is changing the list to:
>
> head -next-> evsel("slots") -next-> evsel("topdown-fe-bound")
>
> which isn't possible in __perf_evlist__set_leader as the head has
> gone. I felt having the list_move in parse_events__set_leader made it
> look like the leader is being placed second rather than at the head.
> So this was an attempt to make the code more intention revealing, but
> that could also be done by some comments, it's just in
> parse_events__set_leader we have the two styles of lists in play.
>
> While I'm complaining, there are many brittle assumptions about evlist
> that it may be worth adding some more assertions for. There is already
> an assertion that the leader was opened before its siblings. There are
> assumptions that the idx given to an evsel are incremental and that
> for the leader 'idx - prev->idx' will yield the number of list
> elements. My concern is for libperf users, it doesn't seem
> unreasonable that they will want to manipulate the evlist but given
> the assumptions they will likely find bugs - such as idx being used to
> index an array and reordering breaking things. The metric code would
> be a customer of reordering were it to need to move slots, but when I
> experimented with this it easily broke things - which motivates doing
> it in parse-events or more likely for the metrics, when generating the
> parse-events string.
>
> Thanks,
> Ian
>
>
>
>
> > >       free($1);
> > >       $$ = list;
> > >  }
> > >  |
> > >  '{' events '}'
> > >  {
> > > -     struct list_head *list = $2;
> > > +     struct list_head *new_head, *list = $2;
> > >
> > >       inc_group_count(list, _parse_state);
> > > -     parse_events__set_leader(NULL, list, _parse_state);
> > > +     new_head = parse_events__set_leader(NULL, list, _parse_state);
> > > +     if (new_head)
> > > +             list_move(new_head, list);
> >
> > same
> >
> > thanks,
> > jirka
> >

Powered by blists - more mailing lists