tools/perf/util/bpf_counter.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /* Copyright (c) 2019 Facebook */
   4
   5 #include <assert.h>
   6 #include <limits.h>
   7 #include <unistd.h>
   8 #include <sys/file.h>
   9 #include <sys/time.h>
  10 #include <linux/err.h>
  11 #include <linux/zalloc.h>
  12 #include <api/fs/fs.h>
  13 #include <perf/bpf_perf.h>
  14
  15 #include "bpf_counter.h"
  16 #include "bpf-utils.h"
  17 #include "counts.h"
  18 #include "debug.h"
  19 #include "evsel.h"
  20 #include "evlist.h"
  21 #include "target.h"
  22 #include "cgroup.h"
  23 #include "cpumap.h"
  24 #include "thread_map.h"
  25
  26 #include "bpf_skel/bpf_prog_profiler.skel.h"
  27 #include "bpf_skel/bperf_u.h"
  28 #include "bpf_skel/bperf_leader.skel.h"
  29 #include "bpf_skel/bperf_follower.skel.h"
  30
  31 #define ATTR_MAP_SIZE 16
  32
  33 static inline void *u64_to_ptr(__u64 ptr)
  34 {
  35         return (void *)(unsigned long)ptr;
  36 }
  37
  38 static struct bpf_counter *bpf_counter_alloc(void)
  39 {
  40         struct bpf_counter *counter;
  41
  42         counter = zalloc(sizeof(*counter));
  43         if (counter)
  44                 INIT_LIST_HEAD(&counter->list);
  45         return counter;
  46 }
  47
  48 static int bpf_program_profiler__destroy(struct evsel *evsel)
  49 {
  50         struct bpf_counter *counter, *tmp;
  51
  52         list_for_each_entry_safe(counter, tmp,
  53                                  &evsel->bpf_counter_list, list) {
  54                 list_del_init(&counter->list);
  55                 bpf_prog_profiler_bpf__destroy(counter->skel);
  56                 free(counter);
  57         }
  58         assert(list_empty(&evsel->bpf_counter_list));
  59
  60         return 0;
  61 }
  62
  63 static char *bpf_target_prog_name(int tgt_fd)
  64 {
  65         struct bpf_func_info *func_info;
  66         struct perf_bpil *info_linear;
  67         const struct btf_type *t;
  68         struct btf *btf = NULL;
  69         char *name = NULL;
  70
  71         info_linear = get_bpf_prog_info_linear(tgt_fd, 1UL << PERF_BPIL_FUNC_INFO);
  72         if (IS_ERR_OR_NULL(info_linear)) {
  73                 pr_debug("failed to get info_linear for prog FD %d\n", tgt_fd);
  74                 return NULL;
  75         }
  76
  77         if (info_linear->info.btf_id == 0) {
  78                 pr_debug("prog FD %d doesn't have valid btf\n", tgt_fd);
  79                 goto out;
  80         }
  81
  82         btf = btf__load_from_kernel_by_id(info_linear->info.btf_id);
  83         if (libbpf_get_error(btf)) {
  84                 pr_debug("failed to load btf for prog FD %d\n", tgt_fd);
  85                 goto out;
  86         }
  87
  88         func_info = u64_to_ptr(info_linear->info.func_info);
  89         t = btf__type_by_id(btf, func_info[0].type_id);
  90         if (!t) {
  91                 pr_debug("btf %d doesn't have type %d\n",
  92                          info_linear->info.btf_id, func_info[0].type_id);
  93                 goto out;
  94         }
  95         name = strdup(btf__name_by_offset(btf, t->name_off));
  96 out:
  97         btf__free(btf);
  98         free(info_linear);
  99         return name;
 100 }
 101
 102 static int bpf_program_profiler_load_one(struct evsel *evsel, u32 prog_id)
 103 {
 104         struct bpf_prog_profiler_bpf *skel;
 105         struct bpf_counter *counter;
 106         struct bpf_program *prog;
 107         char *prog_name;
 108         int prog_fd;
 109         int err;
 110
 111         prog_fd = bpf_prog_get_fd_by_id(prog_id);
 112         if (prog_fd < 0) {
 113                 pr_err("Failed to open fd for bpf prog %u\n", prog_id);
 114                 return -1;
 115         }
 116         counter = bpf_counter_alloc();
 117         if (!counter) {
 118                 close(prog_fd);
 119                 return -1;
 120         }
 121
 122         skel = bpf_prog_profiler_bpf__open();
 123         if (!skel) {
 124                 pr_err("Failed to open bpf skeleton\n");
 125                 goto err_out;
 126         }
 127
 128         skel->rodata->num_cpu = evsel__nr_cpus(evsel);
 129
 130         bpf_map__set_max_entries(skel->maps.events, evsel__nr_cpus(evsel));
 131         bpf_map__set_max_entries(skel->maps.fentry_readings, 1);
 132         bpf_map__set_max_entries(skel->maps.accum_readings, 1);
 133
 134         prog_name = bpf_target_prog_name(prog_fd);
 135         if (!prog_name) {
 136                 pr_err("Failed to get program name for bpf prog %u. Does it have BTF?\n", prog_id);
 137                 goto err_out;
 138         }
 139
 140         bpf_object__for_each_program(prog, skel->obj) {
 141                 err = bpf_program__set_attach_target(prog, prog_fd, prog_name);
 142                 if (err) {
 143                         pr_err("bpf_program__set_attach_target failed.\n"
 144                                "Does bpf prog %u have BTF?\n", prog_id);
 145                         goto err_out;
 146                 }
 147         }
 148         set_max_rlimit();
 149         err = bpf_prog_profiler_bpf__load(skel);
 150         if (err) {
 151                 pr_err("bpf_prog_profiler_bpf__load failed\n");
 152                 goto err_out;
 153         }
 154
 155         assert(skel != NULL);
 156         counter->skel = skel;
 157         list_add(&counter->list, &evsel->bpf_counter_list);
 158         close(prog_fd);
 159         return 0;
 160 err_out:
 161         bpf_prog_profiler_bpf__destroy(skel);
 162         free(counter);
 163         close(prog_fd);
 164         return -1;
 165 }
 166
 167 static int bpf_program_profiler__load(struct evsel *evsel, struct target *target)
 168 {
 169         char *bpf_str, *bpf_str_, *tok, *saveptr = NULL, *p;
 170         u32 prog_id;
 171         int ret;
 172
 173         bpf_str_ = bpf_str = strdup(target->bpf_str);
 174         if (!bpf_str)
 175                 return -1;
 176
 177         while ((tok = strtok_r(bpf_str, ",", &saveptr)) != NULL) {
 178                 prog_id = strtoul(tok, &p, 10);
 179                 if (prog_id == 0 || prog_id == UINT_MAX ||
 180                     (*p != '\0' && *p != ',')) {
 181                         pr_err("Failed to parse bpf prog ids %s\n",
 182                                target->bpf_str);
 183                         return -1;
 184                 }
 185
 186                 ret = bpf_program_profiler_load_one(evsel, prog_id);
 187                 if (ret) {
 188                         bpf_program_profiler__destroy(evsel);
 189                         free(bpf_str_);
 190                         return -1;
 191                 }
 192                 bpf_str = NULL;
 193         }
 194         free(bpf_str_);
 195         return 0;
 196 }
 197
 198 static int bpf_program_profiler__enable(struct evsel *evsel)
 199 {
 200         struct bpf_counter *counter;
 201         int ret;
 202
 203         list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
 204                 assert(counter->skel != NULL);
 205                 ret = bpf_prog_profiler_bpf__attach(counter->skel);
 206                 if (ret) {
 207                         bpf_program_profiler__destroy(evsel);
 208                         return ret;
 209                 }
 210         }
 211         return 0;
 212 }
 213
 214 static int bpf_program_profiler__disable(struct evsel *evsel)
 215 {
 216         struct bpf_counter *counter;
 217
 218         list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
 219                 assert(counter->skel != NULL);
 220                 bpf_prog_profiler_bpf__detach(counter->skel);
 221         }
 222         return 0;
 223 }
 224
 225 static int bpf_program_profiler__read(struct evsel *evsel)
 226 {
 227         // BPF_MAP_TYPE_PERCPU_ARRAY uses /sys/devices/system/cpu/possible
 228         // Sometimes possible > online, like on a Ryzen 3900X that has 24
 229         // threads but its possible showed 0-31 -acme
 230         int num_cpu_bpf = libbpf_num_possible_cpus();
 231         struct bpf_perf_event_value values[num_cpu_bpf];
 232         struct bpf_counter *counter;
 233         struct perf_counts_values *counts;
 234         int reading_map_fd;
 235         __u32 key = 0;
 236         int err, idx, bpf_cpu;
 237
 238         if (list_empty(&evsel->bpf_counter_list))
 239                 return -EAGAIN;
 240
 241         perf_cpu_map__for_each_idx(idx, evsel__cpus(evsel)) {
 242                 counts = perf_counts(evsel->counts, idx, 0);
 243                 counts->val = 0;
 244                 counts->ena = 0;
 245                 counts->run = 0;
 246         }
 247         list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
 248                 struct bpf_prog_profiler_bpf *skel = counter->skel;
 249
 250                 assert(skel != NULL);
 251                 reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
 252
 253                 err = bpf_map_lookup_elem(reading_map_fd, &key, values);
 254                 if (err) {
 255                         pr_err("failed to read value\n");
 256                         return err;
 257                 }
 258
 259                 for (bpf_cpu = 0; bpf_cpu < num_cpu_bpf; bpf_cpu++) {
 260                         idx = perf_cpu_map__idx(evsel__cpus(evsel),
 261                                                 (struct perf_cpu){.cpu = bpf_cpu});
 262                         if (idx == -1)
 263                                 continue;
 264                         counts = perf_counts(evsel->counts, idx, 0);
 265                         counts->val += values[bpf_cpu].counter;
 266                         counts->ena += values[bpf_cpu].enabled;
 267                         counts->run += values[bpf_cpu].running;
 268                 }
 269         }
 270         return 0;
 271 }
 272
 273 static int bpf_program_profiler__install_pe(struct evsel *evsel, int cpu_map_idx,
 274                                             int fd)
 275 {
 276         struct bpf_prog_profiler_bpf *skel;
 277         struct bpf_counter *counter;
 278         int ret;
 279
 280         list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
 281                 skel = counter->skel;
 282                 assert(skel != NULL);
 283
 284                 ret = bpf_map_update_elem(bpf_map__fd(skel->maps.events),
 285                                           &cpu_map_idx, &fd, BPF_ANY);
 286                 if (ret)
 287                         return ret;
 288         }
 289         return 0;
 290 }
 291
 292 struct bpf_counter_ops bpf_program_profiler_ops = {
 293         .load       = bpf_program_profiler__load,
 294         .enable     = bpf_program_profiler__enable,
 295         .disable    = bpf_program_profiler__disable,
 296         .read       = bpf_program_profiler__read,
 297         .destroy    = bpf_program_profiler__destroy,
 298         .install_pe = bpf_program_profiler__install_pe,
 299 };
 300
 301 static bool bperf_attr_map_compatible(int attr_map_fd)
 302 {
 303         struct bpf_map_info map_info = {0};
 304         __u32 map_info_len = sizeof(map_info);
 305         int err;
 306
 307         err = bpf_obj_get_info_by_fd(attr_map_fd, &map_info, &map_info_len);
 308
 309         if (err)
 310                 return false;
 311         return (map_info.key_size == sizeof(struct perf_event_attr)) &&
 312                 (map_info.value_size == sizeof(struct perf_event_attr_map_entry));
 313 }
 314
 315 #ifndef HAVE_LIBBPF_BPF_MAP_CREATE
 316 LIBBPF_API int bpf_create_map(enum bpf_map_type map_type, int key_size,
 317                               int value_size, int max_entries, __u32 map_flags);
 318 int
 319 bpf_map_create(enum bpf_map_type map_type,
 320                const char *map_name __maybe_unused,
 321                __u32 key_size,
 322                __u32 value_size,
 323                __u32 max_entries,
 324                const struct bpf_map_create_opts *opts __maybe_unused)
 325 {
 326 #pragma GCC diagnostic push
 327 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 328         return bpf_create_map(map_type, key_size, value_size, max_entries, 0);
 329 #pragma GCC diagnostic pop
 330 }
 331 #endif
 332
 333 static int bperf_lock_attr_map(struct target *target)
 334 {
 335         char path[PATH_MAX];
 336         int map_fd, err;
 337
 338         if (target->attr_map) {
 339                 scnprintf(path, PATH_MAX, "%s", target->attr_map);
 340         } else {
 341                 scnprintf(path, PATH_MAX, "%s/fs/bpf/%s", sysfs__mountpoint(),
 342                           BPF_PERF_DEFAULT_ATTR_MAP_PATH);
 343         }
 344
 345         if (access(path, F_OK)) {
 346                 map_fd = bpf_map_create(BPF_MAP_TYPE_HASH, NULL,
 347                                         sizeof(struct perf_event_attr),
 348                                         sizeof(struct perf_event_attr_map_entry),
 349                                         ATTR_MAP_SIZE, NULL);
 350                 if (map_fd < 0)
 351                         return -1;
 352
 353                 err = bpf_obj_pin(map_fd, path);
 354                 if (err) {
 355                         /* someone pinned the map in parallel? */
 356                         close(map_fd);
 357                         map_fd = bpf_obj_get(path);
 358                         if (map_fd < 0)
 359                                 return -1;
 360                 }
 361         } else {
 362                 map_fd = bpf_obj_get(path);
 363                 if (map_fd < 0)
 364                         return -1;
 365         }
 366
 367         if (!bperf_attr_map_compatible(map_fd)) {
 368                 close(map_fd);
 369                 return -1;
 370
 371         }
 372         err = flock(map_fd, LOCK_EX);
 373         if (err) {
 374                 close(map_fd);
 375                 return -1;
 376         }
 377         return map_fd;
 378 }
 379
 380 static int bperf_check_target(struct evsel *evsel,
 381                               struct target *target,
 382                               enum bperf_filter_type *filter_type,
 383                               __u32 *filter_entry_cnt)
 384 {
 385         if (evsel->core.leader->nr_members > 1) {
 386                 pr_err("bpf managed perf events do not yet support groups.\n");
 387                 return -1;
 388         }
 389
 390         /* determine filter type based on target */
 391         if (target->system_wide) {
 392                 *filter_type = BPERF_FILTER_GLOBAL;
 393                 *filter_entry_cnt = 1;
 394         } else if (target->cpu_list) {
 395                 *filter_type = BPERF_FILTER_CPU;
 396                 *filter_entry_cnt = perf_cpu_map__nr(evsel__cpus(evsel));
 397         } else if (target->tid) {
 398                 *filter_type = BPERF_FILTER_PID;
 399                 *filter_entry_cnt = perf_thread_map__nr(evsel->core.threads);
 400         } else if (target->pid || evsel->evlist->workload.pid != -1) {
 401                 *filter_type = BPERF_FILTER_TGID;
 402                 *filter_entry_cnt = perf_thread_map__nr(evsel->core.threads);
 403         } else {
 404                 pr_err("bpf managed perf events do not yet support these targets.\n");
 405                 return -1;
 406         }
 407
 408         return 0;
 409 }
 410
 411 static  struct perf_cpu_map *all_cpu_map;
 412
 413 static int bperf_reload_leader_program(struct evsel *evsel, int attr_map_fd,
 414                                        struct perf_event_attr_map_entry *entry)
 415 {
 416         struct bperf_leader_bpf *skel = bperf_leader_bpf__open();
 417         int link_fd, diff_map_fd, err;
 418         struct bpf_link *link = NULL;
 419
 420         if (!skel) {
 421                 pr_err("Failed to open leader skeleton\n");
 422                 return -1;
 423         }
 424
 425         bpf_map__set_max_entries(skel->maps.events, libbpf_num_possible_cpus());
 426         err = bperf_leader_bpf__load(skel);
 427         if (err) {
 428                 pr_err("Failed to load leader skeleton\n");
 429                 goto out;
 430         }
 431
 432         link = bpf_program__attach(skel->progs.on_switch);
 433         if (IS_ERR(link)) {
 434                 pr_err("Failed to attach leader program\n");
 435                 err = PTR_ERR(link);
 436                 goto out;
 437         }
 438
 439         link_fd = bpf_link__fd(link);
 440         diff_map_fd = bpf_map__fd(skel->maps.diff_readings);
 441         entry->link_id = bpf_link_get_id(link_fd);
 442         entry->diff_map_id = bpf_map_get_id(diff_map_fd);
 443         err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, entry, BPF_ANY);
 444         assert(err == 0);
 445
 446         evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry->link_id);
 447         assert(evsel->bperf_leader_link_fd >= 0);
 448
 449         /*
 450          * save leader_skel for install_pe, which is called within
 451          * following evsel__open_per_cpu call
 452          */
 453         evsel->leader_skel = skel;
 454         evsel__open_per_cpu(evsel, all_cpu_map, -1);
 455
 456 out:
 457         bperf_leader_bpf__destroy(skel);
 458         bpf_link__destroy(link);
 459         return err;
 460 }
 461
 462 static int bperf__load(struct evsel *evsel, struct target *target)
 463 {
 464         struct perf_event_attr_map_entry entry = {0xffffffff, 0xffffffff};
 465         int attr_map_fd, diff_map_fd = -1, err;
 466         enum bperf_filter_type filter_type;
 467         __u32 filter_entry_cnt, i;
 468
 469         if (bperf_check_target(evsel, target, &filter_type, &filter_entry_cnt))
 470                 return -1;
 471
 472         if (!all_cpu_map) {
 473                 all_cpu_map = perf_cpu_map__new(NULL);
 474                 if (!all_cpu_map)
 475                         return -1;
 476         }
 477
 478         evsel->bperf_leader_prog_fd = -1;
 479         evsel->bperf_leader_link_fd = -1;
 480
 481         /*
 482          * Step 1: hold a fd on the leader program and the bpf_link, if
 483          * the program is not already gone, reload the program.
 484          * Use flock() to ensure exclusive access to the perf_event_attr
 485          * map.
 486          */
 487         attr_map_fd = bperf_lock_attr_map(target);
 488         if (attr_map_fd < 0) {
 489                 pr_err("Failed to lock perf_event_attr map\n");
 490                 return -1;
 491         }
 492
 493         err = bpf_map_lookup_elem(attr_map_fd, &evsel->core.attr, &entry);
 494         if (err) {
 495                 err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, &entry, BPF_ANY);
 496                 if (err)
 497                         goto out;
 498         }
 499
 500         evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry.link_id);
 501         if (evsel->bperf_leader_link_fd < 0 &&
 502             bperf_reload_leader_program(evsel, attr_map_fd, &entry)) {
 503                 err = -1;
 504                 goto out;
 505         }
 506         /*
 507          * The bpf_link holds reference to the leader program, and the
 508          * leader program holds reference to the maps. Therefore, if
 509          * link_id is valid, diff_map_id should also be valid.
 510          */
 511         evsel->bperf_leader_prog_fd = bpf_prog_get_fd_by_id(
 512                 bpf_link_get_prog_id(evsel->bperf_leader_link_fd));
 513         assert(evsel->bperf_leader_prog_fd >= 0);
 514
 515         diff_map_fd = bpf_map_get_fd_by_id(entry.diff_map_id);
 516         assert(diff_map_fd >= 0);
 517
 518         /*
 519          * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check
 520          * whether the kernel support it
 521          */
 522         err = bperf_trigger_reading(evsel->bperf_leader_prog_fd, 0);
 523         if (err) {
 524                 pr_err("The kernel does not support test_run for raw_tp BPF programs.\n"
 525                        "Therefore, --use-bpf might show inaccurate readings\n");
 526                 goto out;
 527         }
 528
 529         /* Step 2: load the follower skeleton */
 530         evsel->follower_skel = bperf_follower_bpf__open();
 531         if (!evsel->follower_skel) {
 532                 err = -1;
 533                 pr_err("Failed to open follower skeleton\n");
 534                 goto out;
 535         }
 536
 537         /* attach fexit program to the leader program */
 538         bpf_program__set_attach_target(evsel->follower_skel->progs.fexit_XXX,
 539                                        evsel->bperf_leader_prog_fd, "on_switch");
 540
 541         /* connect to leader diff_reading map */
 542         bpf_map__reuse_fd(evsel->follower_skel->maps.diff_readings, diff_map_fd);
 543
 544         /* set up reading map */
 545         bpf_map__set_max_entries(evsel->follower_skel->maps.accum_readings,
 546                                  filter_entry_cnt);
 547         /* set up follower filter based on target */
 548         bpf_map__set_max_entries(evsel->follower_skel->maps.filter,
 549                                  filter_entry_cnt);
 550         err = bperf_follower_bpf__load(evsel->follower_skel);
 551         if (err) {
 552                 pr_err("Failed to load follower skeleton\n");
 553                 bperf_follower_bpf__destroy(evsel->follower_skel);
 554                 evsel->follower_skel = NULL;
 555                 goto out;
 556         }
 557
 558         for (i = 0; i < filter_entry_cnt; i++) {
 559                 int filter_map_fd;
 560                 __u32 key;
 561
 562                 if (filter_type == BPERF_FILTER_PID ||
 563                     filter_type == BPERF_FILTER_TGID)
 564                         key = evsel->core.threads->map[i].pid;
 565                 else if (filter_type == BPERF_FILTER_CPU)
 566                         key = evsel->core.cpus->map[i].cpu;
 567                 else
 568                         break;
 569
 570                 filter_map_fd = bpf_map__fd(evsel->follower_skel->maps.filter);
 571                 bpf_map_update_elem(filter_map_fd, &key, &i, BPF_ANY);
 572         }
 573
 574         evsel->follower_skel->bss->type = filter_type;
 575
 576         err = bperf_follower_bpf__attach(evsel->follower_skel);
 577
 578 out:
 579         if (err && evsel->bperf_leader_link_fd >= 0)
 580                 close(evsel->bperf_leader_link_fd);
 581         if (err && evsel->bperf_leader_prog_fd >= 0)
 582                 close(evsel->bperf_leader_prog_fd);
 583         if (diff_map_fd >= 0)
 584                 close(diff_map_fd);
 585
 586         flock(attr_map_fd, LOCK_UN);
 587         close(attr_map_fd);
 588
 589         return err;
 590 }
 591
 592 static int bperf__install_pe(struct evsel *evsel, int cpu_map_idx, int fd)
 593 {
 594         struct bperf_leader_bpf *skel = evsel->leader_skel;
 595
 596         return bpf_map_update_elem(bpf_map__fd(skel->maps.events),
 597                                    &cpu_map_idx, &fd, BPF_ANY);
 598 }
 599
 600 /*
 601  * trigger the leader prog on each cpu, so the accum_reading map could get
 602  * the latest readings.
 603  */
 604 static int bperf_sync_counters(struct evsel *evsel)
 605 {
 606         int num_cpu, i, cpu;
 607
 608         num_cpu = all_cpu_map->nr;
 609         for (i = 0; i < num_cpu; i++) {
 610                 cpu = all_cpu_map->map[i].cpu;
 611                 bperf_trigger_reading(evsel->bperf_leader_prog_fd, cpu);
 612         }
 613         return 0;
 614 }
 615
 616 static int bperf__enable(struct evsel *evsel)
 617 {
 618         evsel->follower_skel->bss->enabled = 1;
 619         return 0;
 620 }
 621
 622 static int bperf__disable(struct evsel *evsel)
 623 {
 624         evsel->follower_skel->bss->enabled = 0;
 625         return 0;
 626 }
 627
 628 static int bperf__read(struct evsel *evsel)
 629 {
 630         struct bperf_follower_bpf *skel = evsel->follower_skel;
 631         __u32 num_cpu_bpf = cpu__max_cpu().cpu;
 632         struct bpf_perf_event_value values[num_cpu_bpf];
 633         struct perf_counts_values *counts;
 634         int reading_map_fd, err = 0;
 635         __u32 i;
 636         int j;
 637
 638         bperf_sync_counters(evsel);
 639         reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
 640
 641         for (i = 0; i < bpf_map__max_entries(skel->maps.accum_readings); i++) {
 642                 struct perf_cpu entry;
 643                 __u32 cpu;
 644
 645                 err = bpf_map_lookup_elem(reading_map_fd, &i, values);
 646                 if (err)
 647                         goto out;
 648                 switch (evsel->follower_skel->bss->type) {
 649                 case BPERF_FILTER_GLOBAL:
 650                         assert(i == 0);
 651
 652                         perf_cpu_map__for_each_cpu(entry, j, evsel__cpus(evsel)) {
 653                                 counts = perf_counts(evsel->counts, j, 0);
 654                                 counts->val = values[entry.cpu].counter;
 655                                 counts->ena = values[entry.cpu].enabled;
 656                                 counts->run = values[entry.cpu].running;
 657                         }
 658                         break;
 659                 case BPERF_FILTER_CPU:
 660                         cpu = perf_cpu_map__cpu(evsel__cpus(evsel), i).cpu;
 661                         assert(cpu >= 0);
 662                         counts = perf_counts(evsel->counts, i, 0);
 663                         counts->val = values[cpu].counter;
 664                         counts->ena = values[cpu].enabled;
 665                         counts->run = values[cpu].running;
 666                         break;
 667                 case BPERF_FILTER_PID:
 668                 case BPERF_FILTER_TGID:
 669                         counts = perf_counts(evsel->counts, 0, i);
 670                         counts->val = 0;
 671                         counts->ena = 0;
 672                         counts->run = 0;
 673
 674                         for (cpu = 0; cpu < num_cpu_bpf; cpu++) {
 675                                 counts->val += values[cpu].counter;
 676                                 counts->ena += values[cpu].enabled;
 677                                 counts->run += values[cpu].running;
 678                         }
 679                         break;
 680                 default:
 681                         break;
 682                 }
 683         }
 684 out:
 685         return err;
 686 }
 687
 688 static int bperf__destroy(struct evsel *evsel)
 689 {
 690         bperf_follower_bpf__destroy(evsel->follower_skel);
 691         close(evsel->bperf_leader_prog_fd);
 692         close(evsel->bperf_leader_link_fd);
 693         return 0;
 694 }
 695
 696 /*
 697  * bperf: share hardware PMCs with BPF
 698  *
 699  * perf uses performance monitoring counters (PMC) to monitor system
 700  * performance. The PMCs are limited hardware resources. For example,
 701  * Intel CPUs have 3x fixed PMCs and 4x programmable PMCs per cpu.
 702  *
 703  * Modern data center systems use these PMCs in many different ways:
 704  * system level monitoring, (maybe nested) container level monitoring, per
 705  * process monitoring, profiling (in sample mode), etc. In some cases,
 706  * there are more active perf_events than available hardware PMCs. To allow
 707  * all perf_events to have a chance to run, it is necessary to do expensive
 708  * time multiplexing of events.
 709  *
 710  * On the other hand, many monitoring tools count the common metrics
 711  * (cycles, instructions). It is a waste to have multiple tools create
 712  * multiple perf_events of "cycles" and occupy multiple PMCs.
 713  *
 714  * bperf tries to reduce such wastes by allowing multiple perf_events of
 715  * "cycles" or "instructions" (at different scopes) to share PMUs. Instead
 716  * of having each perf-stat session to read its own perf_events, bperf uses
 717  * BPF programs to read the perf_events and aggregate readings to BPF maps.
 718  * Then, the perf-stat session(s) reads the values from these BPF maps.
 719  *
 720  *                                ||
 721  *       shared progs and maps <- || -> per session progs and maps
 722  *                                ||
 723  *   ---------------              ||
 724  *   | perf_events |              ||
 725  *   ---------------       fexit  ||      -----------------
 726  *          |             --------||----> | follower prog |
 727  *       --------------- /        || ---  -----------------
 728  * cs -> | leader prog |/         ||/        |         |
 729  *   --> ---------------         /||  --------------  ------------------
 730  *  /       |         |         / ||  | filter map |  | accum_readings |
 731  * /  ------------  ------------  ||  --------------  ------------------
 732  * |  | prev map |  | diff map |  ||                        |
 733  * |  ------------  ------------  ||                        |
 734  *  \                             ||                        |
 735  * = \ ==================================================== | ============
 736  *    \                                                    /   user space
 737  *     \                                                  /
 738  *      \                                                /
 739  *    BPF_PROG_TEST_RUN                    BPF_MAP_LOOKUP_ELEM
 740  *        \                                            /
 741  *         \                                          /
 742  *          \------  perf-stat ----------------------/
 743  *
 744  * The figure above shows the architecture of bperf. Note that the figure
 745  * is divided into 3 regions: shared progs and maps (top left), per session
 746  * progs and maps (top right), and user space (bottom).
 747  *
 748  * The leader prog is triggered on each context switch (cs). The leader
 749  * prog reads perf_events and stores the difference (current_reading -
 750  * previous_reading) to the diff map. For the same metric, e.g. "cycles",
 751  * multiple perf-stat sessions share the same leader prog.
 752  *
 753  * Each perf-stat session creates a follower prog as fexit program to the
 754  * leader prog. It is possible to attach up to BPF_MAX_TRAMP_PROGS (38)
 755  * follower progs to the same leader prog. The follower prog checks current
 756  * task and processor ID to decide whether to add the value from the diff
 757  * map to its accumulated reading map (accum_readings).
 758  *
 759  * Finally, perf-stat user space reads the value from accum_reading map.
 760  *
 761  * Besides context switch, it is also necessary to trigger the leader prog
 762  * before perf-stat reads the value. Otherwise, the accum_reading map may
 763  * not have the latest reading from the perf_events. This is achieved by
 764  * triggering the event via sys_bpf(BPF_PROG_TEST_RUN) to each CPU.
 765  *
 766  * Comment before the definition of struct perf_event_attr_map_entry
 767  * describes how different sessions of perf-stat share information about
 768  * the leader prog.
 769  */
 770
 771 struct bpf_counter_ops bperf_ops = {
 772         .load       = bperf__load,
 773         .enable     = bperf__enable,
 774         .disable    = bperf__disable,
 775         .read       = bperf__read,
 776         .install_pe = bperf__install_pe,
 777         .destroy    = bperf__destroy,
 778 };
 779
 780 extern struct bpf_counter_ops bperf_cgrp_ops;
 781
 782 static inline bool bpf_counter_skip(struct evsel *evsel)
 783 {
 784         return list_empty(&evsel->bpf_counter_list) &&
 785                 evsel->follower_skel == NULL;
 786 }
 787
 788 int bpf_counter__install_pe(struct evsel *evsel, int cpu_map_idx, int fd)
 789 {
 790         if (bpf_counter_skip(evsel))
 791                 return 0;
 792         return evsel->bpf_counter_ops->install_pe(evsel, cpu_map_idx, fd);
 793 }
 794
 795 int bpf_counter__load(struct evsel *evsel, struct target *target)
 796 {
 797         if (target->bpf_str)
 798                 evsel->bpf_counter_ops = &bpf_program_profiler_ops;
 799         else if (cgrp_event_expanded && target->use_bpf)
 800                 evsel->bpf_counter_ops = &bperf_cgrp_ops;
 801         else if (target->use_bpf || evsel->bpf_counter ||
 802                  evsel__match_bpf_counter_events(evsel->name))
 803                 evsel->bpf_counter_ops = &bperf_ops;
 804
 805         if (evsel->bpf_counter_ops)
 806                 return evsel->bpf_counter_ops->load(evsel, target);
 807         return 0;
 808 }
 809
 810 int bpf_counter__enable(struct evsel *evsel)
 811 {
 812         if (bpf_counter_skip(evsel))
 813                 return 0;
 814         return evsel->bpf_counter_ops->enable(evsel);
 815 }
 816
 817 int bpf_counter__disable(struct evsel *evsel)
 818 {
 819         if (bpf_counter_skip(evsel))
 820                 return 0;
 821         return evsel->bpf_counter_ops->disable(evsel);
 822 }
 823
 824 int bpf_counter__read(struct evsel *evsel)
 825 {
 826         if (bpf_counter_skip(evsel))
 827                 return -EAGAIN;
 828         return evsel->bpf_counter_ops->read(evsel);
 829 }
 830
 831 void bpf_counter__destroy(struct evsel *evsel)
 832 {
 833         if (bpf_counter_skip(evsel))
 834                 return;
 835         evsel->bpf_counter_ops->destroy(evsel);
 836         evsel->bpf_counter_ops = NULL;
 837 }