Re: [PATCH v2 1/3] perf-stat: introduce bperf, share hardware PMCs with BPF

[Song Liu]

> On Mar 18, 2021, at 6:49 AM, Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
> 
> On Thu, Mar 18, 2021 at 4:22 PM Song Liu <songliubraving@xxxxxx> wrote:
>> 
>> 
>> 
>>> On Mar 17, 2021, at 10:54 PM, Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
>>> 
>> 
>> [...]
>> 
>>>> +
>>>> +static int bperf_reload_leader_program(struct evsel *evsel, int attr_map_fd,
>>>> +                                      struct perf_event_attr_map_entry *entry)
>>>> +{
>>>> +       struct bperf_leader_bpf *skel = bperf_leader_bpf__open();
>>>> +       int link_fd, diff_map_fd, err;
>>>> +       struct bpf_link *link = NULL;
>>>> +
>>>> +       if (!skel) {
>>>> +               pr_err("Failed to open leader skeleton\n");
>>>> +               return -1;
>>>> +       }
>>>> +
>>>> +       bpf_map__resize(skel->maps.events, libbpf_num_possible_cpus());
>>>> +       err = bperf_leader_bpf__load(skel);
>>>> +       if (err) {
>>>> +               pr_err("Failed to load leader skeleton\n");
>>>> +               goto out;
>>>> +       }
>>>> +
>>>> +       err = -1;
>>>> +       link = bpf_program__attach(skel->progs.on_switch);
>>>> +       if (!link) {
>>>> +               pr_err("Failed to attach leader program\n");
>>>> +               goto out;
>>>> +       }
>>>> +
>>>> +       link_fd = bpf_link__fd(link);
>>>> +       diff_map_fd = bpf_map__fd(skel->maps.diff_readings);
>>>> +       entry->link_id = bpf_link_get_id(link_fd);
>>>> +       entry->diff_map_id = bpf_map_get_id(diff_map_fd);
>>>> +       err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, entry, BPF_ANY);
>>>> +       assert(err == 0);
>>>> +
>>>> +       evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry->link_id);
>>>> +       assert(evsel->bperf_leader_link_fd >= 0);
>>> 
>>> Isn't it the same as link_fd?
>> 
>> This is a different fd on the same link.
> 
> Ok
> 
>> 
>>> 
>>>> +
>>>> +       /*
>>>> +        * save leader_skel for install_pe, which is called within
>>>> +        * following evsel__open_per_cpu call
>>>> +        */
>>>> +       evsel->leader_skel = skel;
>>>> +       evsel__open_per_cpu(evsel, all_cpu_map, -1);
>>>> +
>>>> +out:
>>>> +       bperf_leader_bpf__destroy(skel);
>>>> +       bpf_link__destroy(link);
>>> 
>>> Why do we destroy it?  Is it because we get an another reference?
>> 
>> Yes. We only need evsel->bperf_leader_link_fd to keep the whole
>> skeleton attached.
>> 
>> When multiple perf-stat sessions are sharing the leader skeleton,
>> only the first one loads the leader skeleton, by calling
>> bperf_reload_leader_program(). Other sessions simply hold a fd to
>> the bpf_link. More explanation in bperf__load() below.
> 
> Ok.
> 
>> 
>> 
>>> 
>>>> +       return err;
>>>> +}
>>>> +
>>>> +static int bperf__load(struct evsel *evsel, struct target *target)
>>>> +{
>>>> +       struct perf_event_attr_map_entry entry = {0xffffffff, 0xffffffff};
>>>> +       int attr_map_fd, diff_map_fd = -1, err;
>>>> +       enum bperf_filter_type filter_type;
>>>> +       __u32 filter_entry_cnt, i;
>>>> +
>>>> +       if (bperf_check_target(evsel, target, &filter_type, &filter_entry_cnt))
>>>> +               return -1;
>>>> +
>>>> +       if (!all_cpu_map) {
>>>> +               all_cpu_map = perf_cpu_map__new(NULL);
>>>> +               if (!all_cpu_map)
>>>> +                       return -1;
>>>> +       }
>>>> +
>>>> +       evsel->bperf_leader_prog_fd = -1;
>>>> +       evsel->bperf_leader_link_fd = -1;
>>>> +
>>>> +       /*
>>>> +        * Step 1: hold a fd on the leader program and the bpf_link, if
>>>> +        * the program is not already gone, reload the program.
>>>> +        * Use flock() to ensure exclusive access to the perf_event_attr
>>>> +        * map.
>>>> +        */
>>>> +       attr_map_fd = bperf_lock_attr_map(target);
>>>> +       if (attr_map_fd < 0) {
>>>> +               pr_err("Failed to lock perf_event_attr map\n");
>>>> +               return -1;
>>>> +       }
>>>> +
>>>> +       err = bpf_map_lookup_elem(attr_map_fd, &evsel->core.attr, &entry);
>>>> +       if (err) {
>>>> +               err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, &entry, BPF_ANY);
>>>> +               if (err)
>>>> +                       goto out;
>>>> +       }
>>>> +
>>>> +       evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry.link_id);
>>>> +       if (evsel->bperf_leader_link_fd < 0 &&
>>>> +           bperf_reload_leader_program(evsel, attr_map_fd, &entry))
>>>> +               goto out;
>> 
>> Continue with previous explanation. In bperf_reload_leader_program(),
>> we open another reference to the link, and destroy the skeleton. This
>> brings the code to the same state as evsel->bperf_leader_link_fd >=
>> condition above.
> 
> Thanks for the explanation.
> 
>> 
>>>> +
>>>> +       /*
>>>> +        * The bpf_link holds reference to the leader program, and the
>>>> +        * leader program holds reference to the maps. Therefore, if
>>>> +        * link_id is valid, diff_map_id should also be valid.
>>>> +        */
>>>> +       evsel->bperf_leader_prog_fd = bpf_prog_get_fd_by_id(
>>>> +               bpf_link_get_prog_id(evsel->bperf_leader_link_fd));
>>>> +       assert(evsel->bperf_leader_prog_fd >= 0);
>>>> +
>>>> +       diff_map_fd = bpf_map_get_fd_by_id(entry.diff_map_id);
>>>> +       assert(diff_map_fd >= 0);
>>>> +
>> 
>> [...]
>> 
>>>> +static int bperf__read(struct evsel *evsel)
>>>> +{
>>>> +       struct bperf_follower_bpf *skel = evsel->follower_skel;
>>>> +       __u32 num_cpu_bpf = cpu__max_cpu();
>>>> +       struct bpf_perf_event_value values[num_cpu_bpf];
>>>> +       int reading_map_fd, err = 0;
>>>> +       __u32 i, j, num_cpu;
>>>> +
>>>> +       bperf_sync_counters(evsel);
>>>> +       reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
>>>> +
>>>> +       for (i = 0; i < bpf_map__max_entries(skel->maps.accum_readings); i++) {
>>>> +               __u32 cpu;
>>>> +
>>>> +               err = bpf_map_lookup_elem(reading_map_fd, &i, values);
>>>> +               if (err)
>>>> +                       goto out;
>>>> +               switch (evsel->follower_skel->bss->type) {
>>>> +               case BPERF_FILTER_GLOBAL:
>>>> +                       assert(i == 0);
>>>> +
>>>> +                       num_cpu = all_cpu_map->nr;
>>>> +                       for (j = 0; j < num_cpu; j++) {
>>>> +                               cpu = all_cpu_map->map[j];
>>>> +                               perf_counts(evsel->counts, cpu, 0)->val = values[cpu].counter;
>>>> +                               perf_counts(evsel->counts, cpu, 0)->ena = values[cpu].enabled;
>>>> +                               perf_counts(evsel->counts, cpu, 0)->run = values[cpu].running;
>>> 
>>> I'm confused with this.  Does the accum_readings map contain values
>>> for all cpus?  IIUC it has only a single entry but you access it for each cpu.
>>> What am I missing?
>> 
>> accumulated_reading is a percpu array. In this case, each cpu has its own
>> bpf_perf_event_value with index 0. The BPF program could only access the
>> data on current cpu. When reading from use space, we get #-of-cpus entries
>> for index 0.
>> 
>> Does this make sense?
> 
> Yep, I didn't know it returns all values when reading from user space.  Then
> I think per cpu event doesn't have many entries too.  Like the global case
> it can simply put the value with key 0, no?

Current per cpu event use same logic as per task events, so we do have multiple
entries. I think it is possible to modify the logic to use one entry for per
cpu events. 

Thanks,
Song