Quellcode-Bibliothek bpf_counter_cgroup.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0

/* Copyright (c) 2021 Facebook */
/* Copyright (c) 2021 Google */

#include <assert.h>
#include <limits.h>
#include <unistd.h>
#include <sys/file.h>
#include <sys/time.h>
#include <sys/resource.h>
#include <linux/err.h>
#include <linux/zalloc.h>
#include <linux/perf_event.h>
#include <api/fs/fs.h>
#include <perf/bpf_perf.h>

#include "affinity.h"
#include "bpf_counter.h"
#include "cgroup.h"
#include "counts.h"
#include "debug.h"
#include "evsel.h"
#include "evlist.h"
#include "target.h"
#include "cpumap.h"
#include "thread_map.h"

#include "bpf_skel/bperf_cgroup.skel.h"

static struct perf_event_attr cgrp_switch_attr = {
 .type = PERF_TYPE_SOFTWARE,
 .config = PERF_COUNT_SW_CGROUP_SWITCHES,
 .size = sizeof(cgrp_switch_attr),
 .sample_period = 1,
 .disabled = 1,
};

static struct evsel *cgrp_switch;
static struct bperf_cgroup_bpf *skel;

#define FD(evt, cpu) (*(int *)xyarray__entry(evt->core.fd, cpu, 0))

static int bperf_load_program(struct evlist *evlist)
{
 struct bpf_link *link;
 struct evsel *evsel;
 struct cgroup *cgrp, *leader_cgrp;
 int i, j;
 struct perf_cpu cpu;
 int total_cpus = cpu__max_cpu().cpu;
 int map_size, map_fd;
 int prog_fd, err;

 skel = bperf_cgroup_bpf__open();
 if (!skel) {
  pr_err("Failed to open cgroup skeleton\n");
  return -1;
 }

 skel->rodata->num_cpus = total_cpus;
 skel->rodata->num_events = evlist->core.nr_entries / nr_cgroups;

 if (cgroup_is_v2("perf_event") > 0)
  skel->rodata->use_cgroup_v2 = 1;

 BUG_ON(evlist->core.nr_entries % nr_cgroups != 0);

 /* we need one copy of events per cpu for reading */
 map_size = total_cpus * evlist->core.nr_entries / nr_cgroups;
 bpf_map__set_max_entries(skel->maps.events, map_size);
 bpf_map__set_max_entries(skel->maps.cgrp_idx, nr_cgroups);
 /* previous result is saved in a per-cpu array */
 map_size = evlist->core.nr_entries / nr_cgroups;
 bpf_map__set_max_entries(skel->maps.prev_readings, map_size);
 /* cgroup result needs all events (per-cpu) */
 map_size = evlist->core.nr_entries;
 bpf_map__set_max_entries(skel->maps.cgrp_readings, map_size);

 set_max_rlimit();

 err = bperf_cgroup_bpf__load(skel);
 if (err) {
  pr_err("Failed to load cgroup skeleton\n");
  goto out;
 }

 err = -1;

 cgrp_switch = evsel__new(&cgrp_switch_attr);
 if (evsel__open_per_cpu(cgrp_switch, evlist->core.all_cpus, -1) < 0) {
  pr_err("Failed to open cgroup switches event\n");
  goto out;
 }

 perf_cpu_map__for_each_cpu(cpu, i, evlist->core.all_cpus) {
  link = bpf_program__attach_perf_event(skel->progs.on_cgrp_switch,
            FD(cgrp_switch, i));
  if (IS_ERR(link)) {
   pr_err("Failed to attach cgroup program\n");
   err = PTR_ERR(link);
   goto out;
  }
 }

 /*
 * Update cgrp_idx map from cgroup-id to event index.
 */
 cgrp = NULL;
 i = 0;

 evlist__for_each_entry(evlist, evsel) {
  if (cgrp == NULL || evsel->cgrp == leader_cgrp) {
   leader_cgrp = evsel->cgrp;
   evsel->cgrp = NULL;

   /* open single copy of the events w/o cgroup */
   err = evsel__open_per_cpu(evsel, evsel->core.cpus, -1);
   if (err == 0)
    evsel->supported = true;

   map_fd = bpf_map__fd(skel->maps.events);
   perf_cpu_map__for_each_cpu(cpu, j, evsel->core.cpus) {
    int fd = FD(evsel, j);
    __u32 idx = evsel->core.idx * total_cpus + cpu.cpu;

    bpf_map_update_elem(map_fd, &idx, &fd, BPF_ANY);
   }

   evsel->cgrp = leader_cgrp;
  }

  if (evsel->cgrp == cgrp)
   continue;

  cgrp = evsel->cgrp;

  if (read_cgroup_id(cgrp) < 0) {
   pr_debug("Failed to get cgroup id for %s\n", cgrp->name);
   cgrp->id = 0;
  }

  map_fd = bpf_map__fd(skel->maps.cgrp_idx);
  err = bpf_map_update_elem(map_fd, &cgrp->id, &i, BPF_ANY);
  if (err < 0) {
   pr_err("Failed to update cgroup index map\n");
   goto out;
  }

  i++;
 }

 /*
 * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check
 * whether the kernel support it
 */
 prog_fd = bpf_program__fd(skel->progs.trigger_read);
 err = bperf_trigger_reading(prog_fd, 0);
 if (err) {
  pr_warning("The kernel does not support test_run for raw_tp BPF programs.\n"
      "Therefore, --for-each-cgroup might show inaccurate readings\n");
  err = 0;
 }

out:
 return err;
}

static int bperf_cgrp__load(struct evsel *evsel,
       struct target *target __maybe_unused)
{
 static bool bperf_loaded = false;

 evsel->bperf_leader_prog_fd = -1;
 evsel->bperf_leader_link_fd = -1;

 if (!bperf_loaded && bperf_load_program(evsel->evlist))
  return -1;

 bperf_loaded = true;
 /* just to bypass bpf_counter_skip() */
 evsel->follower_skel = (struct bperf_follower_bpf *)skel;

 return 0;
}

static int bperf_cgrp__install_pe(struct evsel *evsel __maybe_unused,
      int cpu_map_idx __maybe_unused,
      int fd __maybe_unused)
{
 /* nothing to do */
 return 0;
}

/*
 * trigger the leader prog on each cpu, so the cgrp_reading map could get
 * the latest results.
 */
static int bperf_cgrp__sync_counters(struct evlist *evlist)
{
 struct perf_cpu cpu;
 int idx;
 int prog_fd = bpf_program__fd(skel->progs.trigger_read);

 perf_cpu_map__for_each_cpu(cpu, idx, evlist->core.all_cpus)
  bperf_trigger_reading(prog_fd, cpu.cpu);

 return 0;
}

static int bperf_cgrp__enable(struct evsel *evsel)
{
 if (evsel->core.idx)
  return 0;

 bperf_cgrp__sync_counters(evsel->evlist);

 skel->bss->enabled = 1;
 return 0;
}

static int bperf_cgrp__disable(struct evsel *evsel)
{
 if (evsel->core.idx)
  return 0;

 bperf_cgrp__sync_counters(evsel->evlist);

 skel->bss->enabled = 0;
 return 0;
}

static int bperf_cgrp__read(struct evsel *evsel)
{
 struct evlist *evlist = evsel->evlist;
 int total_cpus = cpu__max_cpu().cpu;
 struct perf_counts_values *counts;
 struct bpf_perf_event_value *values;
 int reading_map_fd, err = 0;

 if (evsel->core.idx)
  return 0;

 bperf_cgrp__sync_counters(evsel->evlist);

 values = calloc(total_cpus, sizeof(*values));
 if (values == NULL)
  return -ENOMEM;

 reading_map_fd = bpf_map__fd(skel->maps.cgrp_readings);

 evlist__for_each_entry(evlist, evsel) {
  __u32 idx = evsel->core.idx;
  int i;
  struct perf_cpu cpu;

  err = bpf_map_lookup_elem(reading_map_fd, &idx, values);
  if (err) {
   pr_err("bpf map lookup failed: idx=%u, event=%s, cgrp=%s\n",
          idx, evsel__name(evsel), evsel->cgrp->name);
   goto out;
  }

  perf_cpu_map__for_each_cpu(cpu, i, evsel->core.cpus) {
   counts = perf_counts(evsel->counts, i, 0);
   counts->val = values[cpu.cpu].counter;
   counts->ena = values[cpu.cpu].enabled;
   counts->run = values[cpu.cpu].running;
  }
 }

out:
 free(values);
 return err;
}

static int bperf_cgrp__destroy(struct evsel *evsel)
{
 if (evsel->core.idx)
  return 0;

 bperf_cgroup_bpf__destroy(skel);
 evsel__delete(cgrp_switch);  // it'll destroy on_switch progs too

 return 0;
}

struct bpf_counter_ops bperf_cgrp_ops = {
 .load       = bperf_cgrp__load,
 .enable     = bperf_cgrp__enable,
 .disable    = bperf_cgrp__disable,
 .read       = bperf_cgrp__read,
 .install_pe = bperf_cgrp__install_pe,
 .destroy    = bperf_cgrp__destroy,
};