[PATCH 2/5] perf cpumap: More cpu map reuse by merge.

Arnaldo Carvalho de Melo arnaldo.melo at gmail.com
Mon Mar 28 13:56:07 PDT 2022



On March 28, 2022 5:50:21 PM GMT-03:00, Ian Rogers <irogers at google.com> wrote:
>On Mon, Mar 28, 2022 at 1:26 PM Arnaldo Carvalho de Melo
><acme at kernel.org> wrote:
>>
>> Em Sun, Mar 27, 2022 at 11:24:11PM -0700, Ian Rogers escreveu:
>> > perf_cpu_map__merge will reuse one of its arguments if they are equal or
>> > the other argument is NULL. The arguments could be reused if it is known
>> > one set of values is a subset of the other. For example, a map of 0-1
>> > and a map of just 0 when merged yields the map of 0-1. Currently a new
>> > map is created rather than adding a reference count to the original 0-1
>> > map.
>> >
>> > Signed-off-by: Ian Rogers <irogers at google.com>
>> > ---
>> >  tools/lib/perf/cpumap.c | 38 ++++++++++++++++++++++++++++----------
>> >  1 file changed, 28 insertions(+), 10 deletions(-)
>> >
>> > diff --git a/tools/lib/perf/cpumap.c b/tools/lib/perf/cpumap.c
>> > index ee66760f1e63..953bc50b0e41 100644
>> > --- a/tools/lib/perf/cpumap.c
>> > +++ b/tools/lib/perf/cpumap.c
>> > @@ -319,6 +319,29 @@ struct perf_cpu perf_cpu_map__max(struct perf_cpu_map *map)
>> >       return map->nr > 0 ? map->map[map->nr - 1] : result;
>> >  }
>> >
>> > +/** Is 'b' a subset of 'a'. */
>> > +static bool perf_cpu_map__is_subset(const struct perf_cpu_map *a,
>> > +                                 const struct perf_cpu_map *b)
>> > +{
>> > +     int i, j;
>> > +
>> > +     if (a == b || !b)
>> > +             return true;
>> > +     if (!a || b->nr > a->nr)
>> > +             return false;
>> > +     j = 0;
>> > +     for (i = 0; i < a->nr; i++) {
>>
>> Since the kernel bumped the minimum gcc version to one that supports
>> declaring loop variables locally and that perf has been using this since
>> forever:
>>
>> ⬢[acme at toolbox perf]$ grep -r '(int [[:alpha:]] = 0;' tools/perf
>> tools/perf/util/block-info.c:   for (int i = 0; i < nr_hpps; i++)
>> tools/perf/util/block-info.c:   for (int i = 0; i < nr_hpps; i++) {
>> tools/perf/util/block-info.c:   for (int i = 0; i < nr_reps; i++)
>> tools/perf/util/stream.c:       for (int i = 0; i < nr_evsel; i++)
>> tools/perf/util/stream.c:       for (int i = 0; i < nr_evsel; i++) {
>> tools/perf/util/stream.c:       for (int i = 0; i < els->nr_evsel; i++) {
>> tools/perf/util/stream.c:       for (int i = 0; i < es_pair->nr_streams; i++) {
>> tools/perf/util/stream.c:       for (int i = 0; i < es_base->nr_streams; i++) {
>> tools/perf/util/cpumap.c:               for (int j = 0; j < c->nr; j++) {
>> tools/perf/util/mem-events.c:   for (int j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
>> tools/perf/util/header.c:       for (int i = 0; i < ff->ph->env.nr_hybrid_cpc_nodes; i++) {
>> tools/perf/builtin-diff.c:      for (int i = 0; i < num; i++)
>> tools/perf/builtin-diff.c:              for (int i = 0; i < pair->block_info->num; i++) {
>> tools/perf/builtin-stat.c:      for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
>> ⬢[acme at toolbox perf]$
>>
>> And this builds on all my test containers, please use:
>>
>>         for (int i = 0, j = 0; i < a->nr; i++)
>>
>> In this case to make the source code more compact.
>
>Ack. We still need to declare 'j' and it is a bit weird to declare j
>before i. Fwiw, Making.config has the CORE_CFLAGS set to gnu99, but
>declaring in the loop is clearly valid in c99.
>
>> > +             if (a->map[i].cpu > b->map[j].cpu)
>> > +                     return false;
>> > +             if (a->map[i].cpu == b->map[j].cpu) {
>> > +                     j++;
>> > +                     if (j == b->nr)
>> > +                             return true;
>>
>> Ok, as its guaranteed that cpu_maps are ordered.
>>
>> > +             }
>> > +     }
>> > +     return false;
>> > +}
>> > +
>> >  /*
>> >   * Merge two cpumaps
>> >   *
>> > @@ -335,17 +358,12 @@ struct perf_cpu_map *perf_cpu_map__merge(struct perf_cpu_map *orig,
>> >       int i, j, k;
>> >       struct perf_cpu_map *merged;
>> >
>> > -     if (!orig && !other)
>> > -             return NULL;
>> > -     if (!orig) {
>> > -             perf_cpu_map__get(other);
>> > -             return other;
>> > -     }
>> > -     if (!other)
>> > -             return orig;
>> > -     if (orig->nr == other->nr &&
>> > -         !memcmp(orig->map, other->map, orig->nr * sizeof(struct perf_cpu)))
>> > +     if (perf_cpu_map__is_subset(orig, other))
>> >               return orig;
>>
>> Can't we have first the introduction of perf_cpu_map__is_subset() and
>> then another patch that gets the refcount, i.e. the four lines below?
>
>I believe that will fail as it'd be an unused static function warning
>and werror.

I thought that it seemed useful enough not to be a static, even if we don't at first export it, i.e. keep it as internal to libperf

>
>Thanks,
>Ian
>
>> > +     if (perf_cpu_map__is_subset(other, orig)) {
>> > +             perf_cpu_map__put(orig);
>> > +             return perf_cpu_map__get(other);
>> > +     }
>> >
>> >       tmp_len = orig->nr + other->nr;
>> >       tmp_cpus = malloc(tmp_len * sizeof(struct perf_cpu));
>> > --
>> > 2.35.1.1021.g381101b075-goog
>>
>> --
>>
>> - Arnaldo



More information about the linux-arm-kernel mailing list