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 "counts.h"
  17 #include "debug.h"
  18 #include "evsel.h"
  19 #include "evlist.h"
  20 #include "target.h"
  21 #include "cgroup.h"
  22 #include "cpumap.h"
  23 #include "thread_map.h"
  24
  25 #include "bpf_skel/bpf_prog_profiler.skel.h"
  26 #include "bpf_skel/bperf_u.h"
  27 #include "bpf_skel/bperf_leader.skel.h"
  28 #include "bpf_skel/bperf_follower.skel.h"
  29
  30 #define ATTR_MAP_SIZE 16
  31
  32 static inline void *u64_to_ptr(__u64 ptr)
  33 {
  34         return (void *)(unsigned long)ptr;
  35 }
  36
  37 static struct bpf_counter *bpf_counter_alloc(void)
  38 {
  39         struct bpf_counter *counter;
  40
  41         counter = zalloc(sizeof(*counter));
  42         if (counter)
  43                 INIT_LIST_HEAD(&counter->list);
  44         return counter;
  45 }
  46
  47 static int bpf_program_profiler__destroy(struct evsel *evsel)
  48 {
  49         struct bpf_counter *counter, *tmp;
  50
  51         list_for_each_entry_safe(counter, tmp,
  52                                  &evsel->bpf_counter_list, list) {
  53                 list_del_init(&counter->list);
  54                 bpf_prog_profiler_bpf__destroy(counter->skel);
  55                 free(counter);
  56         }
  57         assert(list_empty(&evsel->bpf_counter_list));
  58
  59         return 0;
  60 }
  61
  62 static char *bpf_target_prog_name(int tgt_fd)
  63 {
  64         struct bpf_prog_info_linear *info_linear;
  65         struct bpf_func_info *func_info;
  66         const struct btf_type *t;
  67         struct btf *btf = NULL;
  68         char *name = NULL;
  69
  70         info_linear = bpf_program__get_prog_info_linear(
  71                 tgt_fd, 1UL << BPF_PROG_INFO_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__resize(skel->maps.events, evsel__nr_cpus(evsel));
 131         bpf_map__resize(skel->maps.fentry_readings, 1);
 132         bpf_map__resize(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         // perf_cpu_map uses /sys/devices/system/cpu/online
 228         int num_cpu = evsel__nr_cpus(evsel);
 229         // BPF_MAP_TYPE_PERCPU_ARRAY uses /sys/devices/system/cpu/possible
 230         // Sometimes possible > online, like on a Ryzen 3900X that has 24
 231         // threads but its possible showed 0-31 -acme
 232         int num_cpu_bpf = libbpf_num_possible_cpus();
 233         struct bpf_perf_event_value values[num_cpu_bpf];
 234         struct bpf_counter *counter;
 235         int reading_map_fd;
 236         __u32 key = 0;
 237         int err, cpu;
 238
 239         if (list_empty(&evsel->bpf_counter_list))
 240                 return -EAGAIN;
 241
 242         for (cpu = 0; cpu < num_cpu; cpu++) {
 243                 perf_counts(evsel->counts, cpu, 0)->val = 0;
 244                 perf_counts(evsel->counts, cpu, 0)->ena = 0;
 245                 perf_counts(evsel->counts, cpu, 0)->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 (cpu = 0; cpu < num_cpu; cpu++) {
 260                         perf_counts(evsel->counts, cpu, 0)->val += values[cpu].counter;
 261                         perf_counts(evsel->counts, cpu, 0)->ena += values[cpu].enabled;
 262                         perf_counts(evsel->counts, cpu, 0)->run += values[cpu].running;
 263                 }
 264         }
 265         return 0;
 266 }
 267
 268 static int bpf_program_profiler__install_pe(struct evsel *evsel, int cpu,
 269                                             int fd)
 270 {
 271         struct bpf_prog_profiler_bpf *skel;
 272         struct bpf_counter *counter;
 273         int ret;
 274
 275         list_for_each_entry(counter, &evsel->bpf_counter_list, list) {
 276                 skel = counter->skel;
 277                 assert(skel != NULL);
 278
 279                 ret = bpf_map_update_elem(bpf_map__fd(skel->maps.events),
 280                                           &cpu, &fd, BPF_ANY);
 281                 if (ret)
 282                         return ret;
 283         }
 284         return 0;
 285 }
 286
 287 struct bpf_counter_ops bpf_program_profiler_ops = {
 288         .load       = bpf_program_profiler__load,
 289         .enable     = bpf_program_profiler__enable,
 290         .disable    = bpf_program_profiler__disable,
 291         .read       = bpf_program_profiler__read,
 292         .destroy    = bpf_program_profiler__destroy,
 293         .install_pe = bpf_program_profiler__install_pe,
 294 };
 295
 296 static bool bperf_attr_map_compatible(int attr_map_fd)
 297 {
 298         struct bpf_map_info map_info = {0};
 299         __u32 map_info_len = sizeof(map_info);
 300         int err;
 301
 302         err = bpf_obj_get_info_by_fd(attr_map_fd, &map_info, &map_info_len);
 303
 304         if (err)
 305                 return false;
 306         return (map_info.key_size == sizeof(struct perf_event_attr)) &&
 307                 (map_info.value_size == sizeof(struct perf_event_attr_map_entry));
 308 }
 309
 310 static int bperf_lock_attr_map(struct target *target)
 311 {
 312         char path[PATH_MAX];
 313         int map_fd, err;
 314
 315         if (target->attr_map) {
 316                 scnprintf(path, PATH_MAX, "%s", target->attr_map);
 317         } else {
 318                 scnprintf(path, PATH_MAX, "%s/fs/bpf/%s", sysfs__mountpoint(),
 319                           BPF_PERF_DEFAULT_ATTR_MAP_PATH);
 320         }
 321
 322         if (access(path, F_OK)) {
 323                 map_fd = bpf_create_map(BPF_MAP_TYPE_HASH,
 324                                         sizeof(struct perf_event_attr),
 325                                         sizeof(struct perf_event_attr_map_entry),
 326                                         ATTR_MAP_SIZE, 0);
 327                 if (map_fd < 0)
 328                         return -1;
 329
 330                 err = bpf_obj_pin(map_fd, path);
 331                 if (err) {
 332                         /* someone pinned the map in parallel? */
 333                         close(map_fd);
 334                         map_fd = bpf_obj_get(path);
 335                         if (map_fd < 0)
 336                                 return -1;
 337                 }
 338         } else {
 339                 map_fd = bpf_obj_get(path);
 340                 if (map_fd < 0)
 341                         return -1;
 342         }
 343
 344         if (!bperf_attr_map_compatible(map_fd)) {
 345                 close(map_fd);
 346                 return -1;
 347
 348         }
 349         err = flock(map_fd, LOCK_EX);
 350         if (err) {
 351                 close(map_fd);
 352                 return -1;
 353         }
 354         return map_fd;
 355 }
 356
 357 static int bperf_check_target(struct evsel *evsel,
 358                               struct target *target,
 359                               enum bperf_filter_type *filter_type,
 360                               __u32 *filter_entry_cnt)
 361 {
 362         if (evsel->core.leader->nr_members > 1) {
 363                 pr_err("bpf managed perf events do not yet support groups.\n");
 364                 return -1;
 365         }
 366
 367         /* determine filter type based on target */
 368         if (target->system_wide) {
 369                 *filter_type = BPERF_FILTER_GLOBAL;
 370                 *filter_entry_cnt = 1;
 371         } else if (target->cpu_list) {
 372                 *filter_type = BPERF_FILTER_CPU;
 373                 *filter_entry_cnt = perf_cpu_map__nr(evsel__cpus(evsel));
 374         } else if (target->tid) {
 375                 *filter_type = BPERF_FILTER_PID;
 376                 *filter_entry_cnt = perf_thread_map__nr(evsel->core.threads);
 377         } else if (target->pid || evsel->evlist->workload.pid != -1) {
 378                 *filter_type = BPERF_FILTER_TGID;
 379                 *filter_entry_cnt = perf_thread_map__nr(evsel->core.threads);
 380         } else {
 381                 pr_err("bpf managed perf events do not yet support these targets.\n");
 382                 return -1;
 383         }
 384
 385         return 0;
 386 }
 387
 388 static  struct perf_cpu_map *all_cpu_map;
 389
 390 static int bperf_reload_leader_program(struct evsel *evsel, int attr_map_fd,
 391                                        struct perf_event_attr_map_entry *entry)
 392 {
 393         struct bperf_leader_bpf *skel = bperf_leader_bpf__open();
 394         int link_fd, diff_map_fd, err;
 395         struct bpf_link *link = NULL;
 396
 397         if (!skel) {
 398                 pr_err("Failed to open leader skeleton\n");
 399                 return -1;
 400         }
 401
 402         bpf_map__resize(skel->maps.events, libbpf_num_possible_cpus());
 403         err = bperf_leader_bpf__load(skel);
 404         if (err) {
 405                 pr_err("Failed to load leader skeleton\n");
 406                 goto out;
 407         }
 408
 409         link = bpf_program__attach(skel->progs.on_switch);
 410         if (IS_ERR(link)) {
 411                 pr_err("Failed to attach leader program\n");
 412                 err = PTR_ERR(link);
 413                 goto out;
 414         }
 415
 416         link_fd = bpf_link__fd(link);
 417         diff_map_fd = bpf_map__fd(skel->maps.diff_readings);
 418         entry->link_id = bpf_link_get_id(link_fd);
 419         entry->diff_map_id = bpf_map_get_id(diff_map_fd);
 420         err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, entry, BPF_ANY);
 421         assert(err == 0);
 422
 423         evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry->link_id);
 424         assert(evsel->bperf_leader_link_fd >= 0);
 425
 426         /*
 427          * save leader_skel for install_pe, which is called within
 428          * following evsel__open_per_cpu call
 429          */
 430         evsel->leader_skel = skel;
 431         evsel__open_per_cpu(evsel, all_cpu_map, -1);
 432
 433 out:
 434         bperf_leader_bpf__destroy(skel);
 435         bpf_link__destroy(link);
 436         return err;
 437 }
 438
 439 static int bperf__load(struct evsel *evsel, struct target *target)
 440 {
 441         struct perf_event_attr_map_entry entry = {0xffffffff, 0xffffffff};
 442         int attr_map_fd, diff_map_fd = -1, err;
 443         enum bperf_filter_type filter_type;
 444         __u32 filter_entry_cnt, i;
 445
 446         if (bperf_check_target(evsel, target, &filter_type, &filter_entry_cnt))
 447                 return -1;
 448
 449         if (!all_cpu_map) {
 450                 all_cpu_map = perf_cpu_map__new(NULL);
 451                 if (!all_cpu_map)
 452                         return -1;
 453         }
 454
 455         evsel->bperf_leader_prog_fd = -1;
 456         evsel->bperf_leader_link_fd = -1;
 457
 458         /*
 459          * Step 1: hold a fd on the leader program and the bpf_link, if
 460          * the program is not already gone, reload the program.
 461          * Use flock() to ensure exclusive access to the perf_event_attr
 462          * map.
 463          */
 464         attr_map_fd = bperf_lock_attr_map(target);
 465         if (attr_map_fd < 0) {
 466                 pr_err("Failed to lock perf_event_attr map\n");
 467                 return -1;
 468         }
 469
 470         err = bpf_map_lookup_elem(attr_map_fd, &evsel->core.attr, &entry);
 471         if (err) {
 472                 err = bpf_map_update_elem(attr_map_fd, &evsel->core.attr, &entry, BPF_ANY);
 473                 if (err)
 474                         goto out;
 475         }
 476
 477         evsel->bperf_leader_link_fd = bpf_link_get_fd_by_id(entry.link_id);
 478         if (evsel->bperf_leader_link_fd < 0 &&
 479             bperf_reload_leader_program(evsel, attr_map_fd, &entry)) {
 480                 err = -1;
 481                 goto out;
 482         }
 483         /*
 484          * The bpf_link holds reference to the leader program, and the
 485          * leader program holds reference to the maps. Therefore, if
 486          * link_id is valid, diff_map_id should also be valid.
 487          */
 488         evsel->bperf_leader_prog_fd = bpf_prog_get_fd_by_id(
 489                 bpf_link_get_prog_id(evsel->bperf_leader_link_fd));
 490         assert(evsel->bperf_leader_prog_fd >= 0);
 491
 492         diff_map_fd = bpf_map_get_fd_by_id(entry.diff_map_id);
 493         assert(diff_map_fd >= 0);
 494
 495         /*
 496          * bperf uses BPF_PROG_TEST_RUN to get accurate reading. Check
 497          * whether the kernel support it
 498          */
 499         err = bperf_trigger_reading(evsel->bperf_leader_prog_fd, 0);
 500         if (err) {
 501                 pr_err("The kernel does not support test_run for raw_tp BPF programs.\n"
 502                        "Therefore, --use-bpf might show inaccurate readings\n");
 503                 goto out;
 504         }
 505
 506         /* Step 2: load the follower skeleton */
 507         evsel->follower_skel = bperf_follower_bpf__open();
 508         if (!evsel->follower_skel) {
 509                 err = -1;
 510                 pr_err("Failed to open follower skeleton\n");
 511                 goto out;
 512         }
 513
 514         /* attach fexit program to the leader program */
 515         bpf_program__set_attach_target(evsel->follower_skel->progs.fexit_XXX,
 516                                        evsel->bperf_leader_prog_fd, "on_switch");
 517
 518         /* connect to leader diff_reading map */
 519         bpf_map__reuse_fd(evsel->follower_skel->maps.diff_readings, diff_map_fd);
 520
 521         /* set up reading map */
 522         bpf_map__set_max_entries(evsel->follower_skel->maps.accum_readings,
 523                                  filter_entry_cnt);
 524         /* set up follower filter based on target */
 525         bpf_map__set_max_entries(evsel->follower_skel->maps.filter,
 526                                  filter_entry_cnt);
 527         err = bperf_follower_bpf__load(evsel->follower_skel);
 528         if (err) {
 529                 pr_err("Failed to load follower skeleton\n");
 530                 bperf_follower_bpf__destroy(evsel->follower_skel);
 531                 evsel->follower_skel = NULL;
 532                 goto out;
 533         }
 534
 535         for (i = 0; i < filter_entry_cnt; i++) {
 536                 int filter_map_fd;
 537                 __u32 key;
 538
 539                 if (filter_type == BPERF_FILTER_PID ||
 540                     filter_type == BPERF_FILTER_TGID)
 541                         key = evsel->core.threads->map[i].pid;
 542                 else if (filter_type == BPERF_FILTER_CPU)
 543                         key = evsel->core.cpus->map[i];
 544                 else
 545                         break;
 546
 547                 filter_map_fd = bpf_map__fd(evsel->follower_skel->maps.filter);
 548                 bpf_map_update_elem(filter_map_fd, &key, &i, BPF_ANY);
 549         }
 550
 551         evsel->follower_skel->bss->type = filter_type;
 552
 553         err = bperf_follower_bpf__attach(evsel->follower_skel);
 554
 555 out:
 556         if (err && evsel->bperf_leader_link_fd >= 0)
 557                 close(evsel->bperf_leader_link_fd);
 558         if (err && evsel->bperf_leader_prog_fd >= 0)
 559                 close(evsel->bperf_leader_prog_fd);
 560         if (diff_map_fd >= 0)
 561                 close(diff_map_fd);
 562
 563         flock(attr_map_fd, LOCK_UN);
 564         close(attr_map_fd);
 565
 566         return err;
 567 }
 568
 569 static int bperf__install_pe(struct evsel *evsel, int cpu, int fd)
 570 {
 571         struct bperf_leader_bpf *skel = evsel->leader_skel;
 572
 573         return bpf_map_update_elem(bpf_map__fd(skel->maps.events),
 574                                    &cpu, &fd, BPF_ANY);
 575 }
 576
 577 /*
 578  * trigger the leader prog on each cpu, so the accum_reading map could get
 579  * the latest readings.
 580  */
 581 static int bperf_sync_counters(struct evsel *evsel)
 582 {
 583         int num_cpu, i, cpu;
 584
 585         num_cpu = all_cpu_map->nr;
 586         for (i = 0; i < num_cpu; i++) {
 587                 cpu = all_cpu_map->map[i];
 588                 bperf_trigger_reading(evsel->bperf_leader_prog_fd, cpu);
 589         }
 590         return 0;
 591 }
 592
 593 static int bperf__enable(struct evsel *evsel)
 594 {
 595         evsel->follower_skel->bss->enabled = 1;
 596         return 0;
 597 }
 598
 599 static int bperf__disable(struct evsel *evsel)
 600 {
 601         evsel->follower_skel->bss->enabled = 0;
 602         return 0;
 603 }
 604
 605 static int bperf__read(struct evsel *evsel)
 606 {
 607         struct bperf_follower_bpf *skel = evsel->follower_skel;
 608         __u32 num_cpu_bpf = cpu__max_cpu();
 609         struct bpf_perf_event_value values[num_cpu_bpf];
 610         int reading_map_fd, err = 0;
 611         __u32 i, j, num_cpu;
 612
 613         bperf_sync_counters(evsel);
 614         reading_map_fd = bpf_map__fd(skel->maps.accum_readings);
 615
 616         for (i = 0; i < bpf_map__max_entries(skel->maps.accum_readings); i++) {
 617                 __u32 cpu;
 618
 619                 err = bpf_map_lookup_elem(reading_map_fd, &i, values);
 620                 if (err)
 621                         goto out;
 622                 switch (evsel->follower_skel->bss->type) {
 623                 case BPERF_FILTER_GLOBAL:
 624                         assert(i == 0);
 625
 626                         num_cpu = all_cpu_map->nr;
 627                         for (j = 0; j < num_cpu; j++) {
 628                                 cpu = all_cpu_map->map[j];
 629                                 perf_counts(evsel->counts, cpu, 0)->val = values[cpu].counter;
 630                                 perf_counts(evsel->counts, cpu, 0)->ena = values[cpu].enabled;
 631                                 perf_counts(evsel->counts, cpu, 0)->run = values[cpu].running;
 632                         }
 633                         break;
 634                 case BPERF_FILTER_CPU:
 635                         cpu = evsel->core.cpus->map[i];
 636                         perf_counts(evsel->counts, i, 0)->val = values[cpu].counter;
 637                         perf_counts(evsel->counts, i, 0)->ena = values[cpu].enabled;
 638                         perf_counts(evsel->counts, i, 0)->run = values[cpu].running;
 639                         break;
 640                 case BPERF_FILTER_PID:
 641                 case BPERF_FILTER_TGID:
 642                         perf_counts(evsel->counts, 0, i)->val = 0;
 643                         perf_counts(evsel->counts, 0, i)->ena = 0;
 644                         perf_counts(evsel->counts, 0, i)->run = 0;
 645
 646                         for (cpu = 0; cpu < num_cpu_bpf; cpu++) {
 647                                 perf_counts(evsel->counts, 0, i)->val += values[cpu].counter;
 648                                 perf_counts(evsel->counts, 0, i)->ena += values[cpu].enabled;
 649                                 perf_counts(evsel->counts, 0, i)->run += values[cpu].running;
 650                         }
 651                         break;
 652                 default:
 653                         break;
 654                 }
 655         }
 656 out:
 657         return err;
 658 }
 659
 660 static int bperf__destroy(struct evsel *evsel)
 661 {
 662         bperf_follower_bpf__destroy(evsel->follower_skel);
 663         close(evsel->bperf_leader_prog_fd);
 664         close(evsel->bperf_leader_link_fd);
 665         return 0;
 666 }
 667
 668 /*
 669  * bperf: share hardware PMCs with BPF
 670  *
 671  * perf uses performance monitoring counters (PMC) to monitor system
 672  * performance. The PMCs are limited hardware resources. For example,
 673  * Intel CPUs have 3x fixed PMCs and 4x programmable PMCs per cpu.
 674  *
 675  * Modern data center systems use these PMCs in many different ways:
 676  * system level monitoring, (maybe nested) container level monitoring, per
 677  * process monitoring, profiling (in sample mode), etc. In some cases,
 678  * there are more active perf_events than available hardware PMCs. To allow
 679  * all perf_events to have a chance to run, it is necessary to do expensive
 680  * time multiplexing of events.
 681  *
 682  * On the other hand, many monitoring tools count the common metrics
 683  * (cycles, instructions). It is a waste to have multiple tools create
 684  * multiple perf_events of "cycles" and occupy multiple PMCs.
 685  *
 686  * bperf tries to reduce such wastes by allowing multiple perf_events of
 687  * "cycles" or "instructions" (at different scopes) to share PMUs. Instead
 688  * of having each perf-stat session to read its own perf_events, bperf uses
 689  * BPF programs to read the perf_events and aggregate readings to BPF maps.
 690  * Then, the perf-stat session(s) reads the values from these BPF maps.
 691  *
 692  *                                ||
 693  *       shared progs and maps <- || -> per session progs and maps
 694  *                                ||
 695  *   ---------------              ||
 696  *   | perf_events |              ||
 697  *   ---------------       fexit  ||      -----------------
 698  *          |             --------||----> | follower prog |
 699  *       --------------- /        || ---  -----------------
 700  * cs -> | leader prog |/         ||/        |         |
 701  *   --> ---------------         /||  --------------  ------------------
 702  *  /       |         |         / ||  | filter map |  | accum_readings |
 703  * /  ------------  ------------  ||  --------------  ------------------
 704  * |  | prev map |  | diff map |  ||                        |
 705  * |  ------------  ------------  ||                        |
 706  *  \                             ||                        |
 707  * = \ ==================================================== | ============
 708  *    \                                                    /   user space
 709  *     \                                                  /
 710  *      \                                                /
 711  *    BPF_PROG_TEST_RUN                    BPF_MAP_LOOKUP_ELEM
 712  *        \                                            /
 713  *         \                                          /
 714  *          \------  perf-stat ----------------------/
 715  *
 716  * The figure above shows the architecture of bperf. Note that the figure
 717  * is divided into 3 regions: shared progs and maps (top left), per session
 718  * progs and maps (top right), and user space (bottom).
 719  *
 720  * The leader prog is triggered on each context switch (cs). The leader
 721  * prog reads perf_events and stores the difference (current_reading -
 722  * previous_reading) to the diff map. For the same metric, e.g. "cycles",
 723  * multiple perf-stat sessions share the same leader prog.
 724  *
 725  * Each perf-stat session creates a follower prog as fexit program to the
 726  * leader prog. It is possible to attach up to BPF_MAX_TRAMP_PROGS (38)
 727  * follower progs to the same leader prog. The follower prog checks current
 728  * task and processor ID to decide whether to add the value from the diff
 729  * map to its accumulated reading map (accum_readings).
 730  *
 731  * Finally, perf-stat user space reads the value from accum_reading map.
 732  *
 733  * Besides context switch, it is also necessary to trigger the leader prog
 734  * before perf-stat reads the value. Otherwise, the accum_reading map may
 735  * not have the latest reading from the perf_events. This is achieved by
 736  * triggering the event via sys_bpf(BPF_PROG_TEST_RUN) to each CPU.
 737  *
 738  * Comment before the definition of struct perf_event_attr_map_entry
 739  * describes how different sessions of perf-stat share information about
 740  * the leader prog.
 741  */
 742
 743 struct bpf_counter_ops bperf_ops = {
 744         .load       = bperf__load,
 745         .enable     = bperf__enable,
 746         .disable    = bperf__disable,
 747         .read       = bperf__read,
 748         .install_pe = bperf__install_pe,
 749         .destroy    = bperf__destroy,
 750 };
 751
 752 extern struct bpf_counter_ops bperf_cgrp_ops;
 753
 754 static inline bool bpf_counter_skip(struct evsel *evsel)
 755 {
 756         return list_empty(&evsel->bpf_counter_list) &&
 757                 evsel->follower_skel == NULL;
 758 }
 759
 760 int bpf_counter__install_pe(struct evsel *evsel, int cpu, int fd)
 761 {
 762         if (bpf_counter_skip(evsel))
 763                 return 0;
 764         return evsel->bpf_counter_ops->install_pe(evsel, cpu, fd);
 765 }
 766
 767 int bpf_counter__load(struct evsel *evsel, struct target *target)
 768 {
 769         if (target->bpf_str)
 770                 evsel->bpf_counter_ops = &bpf_program_profiler_ops;
 771         else if (cgrp_event_expanded && target->use_bpf)
 772                 evsel->bpf_counter_ops = &bperf_cgrp_ops;
 773         else if (target->use_bpf || evsel->bpf_counter ||
 774                  evsel__match_bpf_counter_events(evsel->name))
 775                 evsel->bpf_counter_ops = &bperf_ops;
 776
 777         if (evsel->bpf_counter_ops)
 778                 return evsel->bpf_counter_ops->load(evsel, target);
 779         return 0;
 780 }
 781
 782 int bpf_counter__enable(struct evsel *evsel)
 783 {
 784         if (bpf_counter_skip(evsel))
 785                 return 0;
 786         return evsel->bpf_counter_ops->enable(evsel);
 787 }
 788
 789 int bpf_counter__disable(struct evsel *evsel)
 790 {
 791         if (bpf_counter_skip(evsel))
 792                 return 0;
 793         return evsel->bpf_counter_ops->disable(evsel);
 794 }
 795
 796 int bpf_counter__read(struct evsel *evsel)
 797 {
 798         if (bpf_counter_skip(evsel))
 799                 return -EAGAIN;
 800         return evsel->bpf_counter_ops->read(evsel);
 801 }
 802
 803 void bpf_counter__destroy(struct evsel *evsel)
 804 {
 805         if (bpf_counter_skip(evsel))
 806                 return;
 807         evsel->bpf_counter_ops->destroy(evsel);
 808         evsel->bpf_counter_ops = NULL;
 809 }