Re: [PATCH v1 2/2] perf bpf_counter: Fix handling of cpumap fixing hybrid

["Falcon, Thomas" <thomas.falcon@intel.com>]

On Mon, 2025-10-06 at 09:18 -0700, Ian Rogers wrote:
> On Wed, Oct 1, 2025 at 11:12 AM Ian Rogers <irogers@google.com>
> wrote:
> > 
> > Don't open evsels on all CPUs, open them just on the CPUs they
> > support. This avoids opening say an e-core event on a p-core and
> > getting a failure - achieve this by getting rid of the
> > "all_cpu_map".
> > 
> > In install_pe functions don't use the cpu_map_idx as a CPU number,
> > translate the cpu_map_idx, which is a dense index into the cpu_map
> > skipping holes at the beginning, to a proper CPU number.
> > 
> > Before:
> > ```
> > $ perf stat --bpf-counters -a -e cycles,instructions -- sleep 1
> > 
> >  Performance counter stats for 'system wide':
> > 
> >    <not supported>      cpu_atom/cycles/
> >        566,270,672      cpu_core/cycles/
> >    <not supported>      cpu_atom/instructions/
> >        572,792,836      cpu_core/instructions/           #    1.01 
> > insn per cycle
> > 
> >        1.001595384 seconds time elapsed
> > ```
> > 
> > After:
> > ```
> > $ perf stat --bpf-counters -a -e cycles,instructions -- sleep 1
> > 
> >  Performance counter stats for 'system wide':
> > 
> >        443,299,201      cpu_atom/cycles/
> >      1,233,919,737      cpu_core/cycles/
> >        213,634,112      cpu_atom/instructions/           #    0.48 
> > insn per cycle
> >      2,758,965,527      cpu_core/instructions/           #    2.24 
> > insn per cycle
> > 
> >        1.001699485 seconds time elapsed
> > ```
> > 
> > Fixes: 7fac83aaf2ee ("perf stat: Introduce 'bperf' to share
> > hardware PMCs with BPF")
> > Signed-off-by: Ian Rogers <irogers@google.com>
> 
> +Thomas Falcon
> 
> I think it'd be nice to get this quite major fix for
> --bpf-counters/bperf for hybrid architectures into v6.18 and stable
> builds. Thomas would it be possible for you to give a Tested-by tag
> using the reproduction in the commit message?
> 
> Thanks,
> Ian
> 

Sorry for missing this. Here's my reproduction on an alder lake

sudo ./perf stat --bpf-counters -a -e cycles,instructions -- sleep 1
...

 Performance counter stats for 'system wide':

       364,715,896      cpu_atom/cycles/                              
       946,331,957      cpu_core/cycles/                              
       169,842,929      cpu_atom/instructions/           #    0.47 
insn per cycle            
     1,338,720,324      cpu_core/instructions/           #    1.41 
insn per cycle            

       1.001667769 seconds time elapsed

It looks like it's in perf-tools-next already but have my tested by for
what it's worth.

Tested-by: Thomas Falcon <thomas.falcon@intel.com>


> > ---
> >  tools/perf/util/bpf_counter.c        | 26 ++++++++++--------------
> > --
> >  tools/perf/util/bpf_counter_cgroup.c |  3 ++-
> >  2 files changed, 12 insertions(+), 17 deletions(-)
> > 
> > diff --git a/tools/perf/util/bpf_counter.c
> > b/tools/perf/util/bpf_counter.c
> > index 1c6cb5ea077e..ca5d01b9017d 100644
> > --- a/tools/perf/util/bpf_counter.c
> > +++ b/tools/perf/util/bpf_counter.c
> > @@ -336,6 +336,7 @@ static int
> > bpf_program_profiler__install_pe(struct evsel *evsel, int
> > cpu_map_idx
> >  {
> >         struct bpf_prog_profiler_bpf *skel;
> >         struct bpf_counter *counter;
> > +       int cpu = perf_cpu_map__cpu(evsel->core.cpus,
> > cpu_map_idx).cpu;
> >         int ret;
> > 
> >         list_for_each_entry(counter, &evsel->bpf_counter_list,
> > list) {
> > @@ -343,7 +344,7 @@ static int
> > bpf_program_profiler__install_pe(struct evsel *evsel, int
> > cpu_map_idx
> >                 assert(skel != NULL);
> > 
> >                 ret = bpf_map_update_elem(bpf_map__fd(skel-
> > >maps.events),
> > -                                         &cpu_map_idx, &fd,
> > BPF_ANY);
> > +                                         &cpu, &fd, BPF_ANY);
> >                 if (ret)
> >                         return ret;
> >         }
> > @@ -451,7 +452,6 @@ static int bperf_check_target(struct evsel
> > *evsel,
> >         return 0;
> >  }
> > 
> > -static struct perf_cpu_map *all_cpu_map;
> >  static __u32 filter_entry_cnt;
> > 
> >  static int bperf_reload_leader_program(struct evsel *evsel, int
> > attr_map_fd,
> > @@ -495,7 +495,7 @@ static int bperf_reload_leader_program(struct
> > evsel *evsel, int attr_map_fd,
> >          * following evsel__open_per_cpu call
> >          */
> >         evsel->leader_skel = skel;
> > -       evsel__open_per_cpu(evsel, all_cpu_map, -1);
> > +       evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
> > 
> >  out:
> >         bperf_leader_bpf__destroy(skel);
> > @@ -533,12 +533,6 @@ static int bperf__load(struct evsel *evsel,
> > struct target *target)
> >         if (bperf_check_target(evsel, target, &filter_type,
> > &filter_entry_cnt))
> >                 return -1;
> > 
> > -       if (!all_cpu_map) {
> > -               all_cpu_map = perf_cpu_map__new_online_cpus();
> > -               if (!all_cpu_map)
> > -                       return -1;
> > -       }
> > -
> >         evsel->bperf_leader_prog_fd = -1;
> >         evsel->bperf_leader_link_fd = -1;
> > 
> > @@ -656,9 +650,10 @@ static int bperf__load(struct evsel *evsel,
> > struct target *target)
> >  static int bperf__install_pe(struct evsel *evsel, int cpu_map_idx,
> > int fd)
> >  {
> >         struct bperf_leader_bpf *skel = evsel->leader_skel;
> > +       int cpu = perf_cpu_map__cpu(evsel->core.cpus,
> > cpu_map_idx).cpu;
> > 
> >         return bpf_map_update_elem(bpf_map__fd(skel->maps.events),
> > -                                  &cpu_map_idx, &fd, BPF_ANY);
> > +                                  &cpu, &fd, BPF_ANY);
> >  }
> > 
> >  /*
> > @@ -667,13 +662,12 @@ static int bperf__install_pe(struct evsel
> > *evsel, int cpu_map_idx, int fd)
> >   */
> >  static int bperf_sync_counters(struct evsel *evsel)
> >  {
> > -       int num_cpu, i, cpu;
> > +       struct perf_cpu cpu;
> > +       int idx;
> > +
> > +       perf_cpu_map__for_each_cpu(cpu, idx, evsel->core.cpus)
> > +               bperf_trigger_reading(evsel->bperf_leader_prog_fd,
> > cpu.cpu);
> > 
> > -       num_cpu = perf_cpu_map__nr(all_cpu_map);
> > -       for (i = 0; i < num_cpu; i++) {
> > -               cpu = perf_cpu_map__cpu(all_cpu_map, i).cpu;
> > -               bperf_trigger_reading(evsel->bperf_leader_prog_fd,
> > cpu);
> > -       }
> >         return 0;
> >  }
> > 
> > diff --git a/tools/perf/util/bpf_counter_cgroup.c
> > b/tools/perf/util/bpf_counter_cgroup.c
> > index ed6a29b106b4..690be3ce3e11 100644
> > --- a/tools/perf/util/bpf_counter_cgroup.c
> > +++ b/tools/perf/util/bpf_counter_cgroup.c
> > @@ -186,7 +186,8 @@ static int bperf_cgrp__load(struct evsel
> > *evsel,
> >  }
> > 
> >  static int bperf_cgrp__install_pe(struct evsel *evsel
> > __maybe_unused,
> > -                                 int cpu __maybe_unused, int fd
> > __maybe_unused)
> > +                                 int cpu_map_idx __maybe_unused,
> > +                                 int fd __maybe_unused)
> >  {
> >         /* nothing to do */
> >         return 0;
> > --
> > 2.51.0.618.g983fd99d29-goog
> >