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