tools/perf/builtin-stat.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * builtin-stat.c
   4  *
   5  * Builtin stat command: Give a precise performance counters summary
   6  * overview about any workload, CPU or specific PID.
   7  *
   8  * Sample output:
   9
  10    $ perf stat ./hackbench 10
  11
  12   Time: 0.118
  13
  14   Performance counter stats for './hackbench 10':
  15
  16        1708.761321 task-clock                #   11.037 CPUs utilized
  17             41,190 context-switches          #    0.024 M/sec
  18              6,735 CPU-migrations            #    0.004 M/sec
  19             17,318 page-faults               #    0.010 M/sec
  20      5,205,202,243 cycles                    #    3.046 GHz
  21      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  22      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  23      2,603,501,247 instructions              #    0.50  insns per cycle
  24                                              #    1.48  stalled cycles per insn
  25        484,357,498 branches                  #  283.455 M/sec
  26          6,388,934 branch-misses             #    1.32% of all branches
  27
  28         0.154822978  seconds time elapsed
  29
  30  *
  31  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <[email protected]>
  32  *
  33  * Improvements and fixes by:
  34  *
  35  *   Arjan van de Ven <[email protected]>
  36  *   Yanmin Zhang <[email protected]>
  37  *   Wu Fengguang <[email protected]>
  38  *   Mike Galbraith <[email protected]>
  39  *   Paul Mackerras <[email protected]>
  40  *   Jaswinder Singh Rajput <[email protected]>
  41  */
  42
  43 #include "builtin.h"
  44 #include "util/cgroup.h"
  45 #include <subcmd/parse-options.h>
  46 #include "util/parse-events.h"
  47 #include "util/pmus.h"
  48 #include "util/pmu.h"
  49 #include "util/tool_pmu.h"
  50 #include "util/event.h"
  51 #include "util/evlist.h"
  52 #include "util/evsel.h"
  53 #include "util/debug.h"
  54 #include "util/color.h"
  55 #include "util/stat.h"
  56 #include "util/header.h"
  57 #include "util/cpumap.h"
  58 #include "util/thread_map.h"
  59 #include "util/counts.h"
  60 #include "util/topdown.h"
  61 #include "util/session.h"
  62 #include "util/tool.h"
  63 #include "util/string2.h"
  64 #include "util/metricgroup.h"
  65 #include "util/synthetic-events.h"
  66 #include "util/target.h"
  67 #include "util/time-utils.h"
  68 #include "util/top.h"
  69 #include "util/affinity.h"
  70 #include "util/pfm.h"
  71 #include "util/bpf_counter.h"
  72 #include "util/iostat.h"
  73 #include "util/util.h"
  74 #include "util/intel-tpebs.h"
  75 #include "asm/bug.h"
  76
  77 #include <linux/time64.h>
  78 #include <linux/zalloc.h>
  79 #include <api/fs/fs.h>
  80 #include <errno.h>
  81 #include <signal.h>
  82 #include <stdlib.h>
  83 #include <sys/prctl.h>
  84 #include <inttypes.h>
  85 #include <locale.h>
  86 #include <math.h>
  87 #include <sys/types.h>
  88 #include <sys/stat.h>
  89 #include <sys/wait.h>
  90 #include <unistd.h>
  91 #include <sys/time.h>
  92 #include <sys/resource.h>
  93 #include <linux/err.h>
  94
  95 #include <linux/ctype.h>
  96 #include <perf/evlist.h>
  97 #include <internal/threadmap.h>
  98
  99 #define DEFAULT_SEPARATOR       " "
 100 #define FREEZE_ON_SMI_PATH      "devices/cpu/freeze_on_smi"
 101
 102 static void print_counters(struct timespec *ts, int argc, const char **argv);
 103
 104 static struct evlist    *evsel_list;
 105 static struct parse_events_option_args parse_events_option_args = {
 106         .evlistp = &evsel_list,
 107 };
 108
 109 static bool all_counters_use_bpf = true;
 110
 111 static struct target target = {
 112         .uid    = UINT_MAX,
 113 };
 114
 115 #define METRIC_ONLY_LEN 20
 116
 117 static volatile sig_atomic_t    child_pid                       = -1;
 118 static int                      detailed_run                    =  0;
 119 static bool                     transaction_run;
 120 static bool                     topdown_run                     = false;
 121 static bool                     smi_cost                        = false;
 122 static bool                     smi_reset                       = false;
 123 static int                      big_num_opt                     =  -1;
 124 static const char               *pre_cmd                        = NULL;
 125 static const char               *post_cmd                       = NULL;
 126 static bool                     sync_run                        = false;
 127 static bool                     forever                         = false;
 128 static bool                     force_metric_only               = false;
 129 static struct timespec          ref_time;
 130 static bool                     append_file;
 131 static bool                     interval_count;
 132 static const char               *output_name;
 133 static int                      output_fd;
 134 static char                     *metrics;
 135
 136 struct perf_stat {
 137         bool                     record;
 138         struct perf_data         data;
 139         struct perf_session     *session;
 140         u64                      bytes_written;
 141         struct perf_tool         tool;
 142         bool                     maps_allocated;
 143         struct perf_cpu_map     *cpus;
 144         struct perf_thread_map *threads;
 145         enum aggr_mode           aggr_mode;
 146         u32                      aggr_level;
 147 };
 148
 149 static struct perf_stat         perf_stat;
 150 #define STAT_RECORD             perf_stat.record
 151
 152 static volatile sig_atomic_t done = 0;
 153
 154 static struct perf_stat_config stat_config = {
 155         .aggr_mode              = AGGR_GLOBAL,
 156         .aggr_level             = MAX_CACHE_LVL + 1,
 157         .scale                  = true,
 158         .unit_width             = 4, /* strlen("unit") */
 159         .run_count              = 1,
 160         .metric_only_len        = METRIC_ONLY_LEN,
 161         .walltime_nsecs_stats   = &walltime_nsecs_stats,
 162         .ru_stats               = &ru_stats,
 163         .big_num                = true,
 164         .ctl_fd                 = -1,
 165         .ctl_fd_ack             = -1,
 166         .iostat_run             = false,
 167 };
 168
 169 /* Options set from the command line. */
 170 struct opt_aggr_mode {
 171         bool node, socket, die, cluster, cache, core, thread, no_aggr;
 172 };
 173
 174 /* Turn command line option into most generic aggregation mode setting. */
 175 static enum aggr_mode opt_aggr_mode_to_aggr_mode(struct opt_aggr_mode *opt_mode)
 176 {
 177         enum aggr_mode mode = AGGR_GLOBAL;
 178
 179         if (opt_mode->node)
 180                 mode = AGGR_NODE;
 181         if (opt_mode->socket)
 182                 mode = AGGR_SOCKET;
 183         if (opt_mode->die)
 184                 mode = AGGR_DIE;
 185         if (opt_mode->cluster)
 186                 mode = AGGR_CLUSTER;
 187         if (opt_mode->cache)
 188                 mode = AGGR_CACHE;
 189         if (opt_mode->core)
 190                 mode = AGGR_CORE;
 191         if (opt_mode->thread)
 192                 mode = AGGR_THREAD;
 193         if (opt_mode->no_aggr)
 194                 mode = AGGR_NONE;
 195         return mode;
 196 }
 197
 198 static void evlist__check_cpu_maps(struct evlist *evlist)
 199 {
 200         struct evsel *evsel, *warned_leader = NULL;
 201
 202         evlist__for_each_entry(evlist, evsel) {
 203                 struct evsel *leader = evsel__leader(evsel);
 204
 205                 /* Check that leader matches cpus with each member. */
 206                 if (leader == evsel)
 207                         continue;
 208                 if (perf_cpu_map__equal(leader->core.cpus, evsel->core.cpus))
 209                         continue;
 210
 211                 /* If there's mismatch disable the group and warn user. */
 212                 if (warned_leader != leader) {
 213                         char buf[200];
 214
 215                         pr_warning("WARNING: grouped events cpus do not match.\n"
 216                                 "Events with CPUs not matching the leader will "
 217                                 "be removed from the group.\n");
 218                         evsel__group_desc(leader, buf, sizeof(buf));
 219                         pr_warning("  %s\n", buf);
 220                         warned_leader = leader;
 221                 }
 222                 if (verbose > 0) {
 223                         char buf[200];
 224
 225                         cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
 226                         pr_warning("     %s: %s\n", leader->name, buf);
 227                         cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
 228                         pr_warning("     %s: %s\n", evsel->name, buf);
 229                 }
 230
 231                 evsel__remove_from_group(evsel, leader);
 232         }
 233 }
 234
 235 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 236                                  struct timespec *b)
 237 {
 238         r->tv_sec = a->tv_sec - b->tv_sec;
 239         if (a->tv_nsec < b->tv_nsec) {
 240                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 241                 r->tv_sec--;
 242         } else {
 243                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 244         }
 245 }
 246
 247 static void perf_stat__reset_stats(void)
 248 {
 249         evlist__reset_stats(evsel_list);
 250         perf_stat__reset_shadow_stats();
 251 }
 252
 253 static int process_synthesized_event(const struct perf_tool *tool __maybe_unused,
 254                                      union perf_event *event,
 255                                      struct perf_sample *sample __maybe_unused,
 256                                      struct machine *machine __maybe_unused)
 257 {
 258         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 259                 pr_err("failed to write perf data, error: %m\n");
 260                 return -1;
 261         }
 262
 263         perf_stat.bytes_written += event->header.size;
 264         return 0;
 265 }
 266
 267 static int write_stat_round_event(u64 tm, u64 type)
 268 {
 269         return perf_event__synthesize_stat_round(NULL, tm, type,
 270                                                  process_synthesized_event,
 271                                                  NULL);
 272 }
 273
 274 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 275         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 276
 277 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
 278
 279 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
 280                                    struct perf_counts_values *count)
 281 {
 282         struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
 283         struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
 284
 285         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 286                                            process_synthesized_event, NULL);
 287 }
 288
 289 static int read_single_counter(struct evsel *counter, int cpu_map_idx, int thread)
 290 {
 291         int err = evsel__read_counter(counter, cpu_map_idx, thread);
 292
 293         /*
 294          * Reading user and system time will fail when the process
 295          * terminates. Use the wait4 values in that case.
 296          */
 297         if (err && cpu_map_idx == 0 &&
 298             (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME ||
 299              evsel__tool_event(counter) == TOOL_PMU__EVENT_SYSTEM_TIME)) {
 300                 u64 val, *start_time;
 301                 struct perf_counts_values *count =
 302                         perf_counts(counter->counts, cpu_map_idx, thread);
 303
 304                 start_time = xyarray__entry(counter->start_times, cpu_map_idx, thread);
 305                 if (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME)
 306                         val = ru_stats.ru_utime_usec_stat.mean;
 307                 else
 308                         val = ru_stats.ru_stime_usec_stat.mean;
 309                 count->ena = count->run = *start_time + val;
 310                 count->val = val;
 311                 return 0;
 312         }
 313         return err;
 314 }
 315
 316 /*
 317  * Read out the results of a single counter:
 318  * do not aggregate counts across CPUs in system-wide mode
 319  */
 320 static int read_counter_cpu(struct evsel *counter, int cpu_map_idx)
 321 {
 322         int nthreads = perf_thread_map__nr(evsel_list->core.threads);
 323         int thread;
 324
 325         if (!counter->supported)
 326                 return -ENOENT;
 327
 328         for (thread = 0; thread < nthreads; thread++) {
 329                 struct perf_counts_values *count;
 330
 331                 count = perf_counts(counter->counts, cpu_map_idx, thread);
 332
 333                 /*
 334                  * The leader's group read loads data into its group members
 335                  * (via evsel__read_counter()) and sets their count->loaded.
 336                  */
 337                 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
 338                     read_single_counter(counter, cpu_map_idx, thread)) {
 339                         counter->counts->scaled = -1;
 340                         perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
 341                         perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
 342                         return -1;
 343                 }
 344
 345                 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
 346
 347                 if (STAT_RECORD) {
 348                         if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
 349                                 pr_err("failed to write stat event\n");
 350                                 return -1;
 351                         }
 352                 }
 353
 354                 if (verbose > 1) {
 355                         fprintf(stat_config.output,
 356                                 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 357                                         evsel__name(counter),
 358                                         perf_cpu_map__cpu(evsel__cpus(counter),
 359                                                           cpu_map_idx).cpu,
 360                                         count->val, count->ena, count->run);
 361                 }
 362         }
 363
 364         return 0;
 365 }
 366
 367 static int read_affinity_counters(void)
 368 {
 369         struct evlist_cpu_iterator evlist_cpu_itr;
 370         struct affinity saved_affinity, *affinity;
 371
 372         if (all_counters_use_bpf)
 373                 return 0;
 374
 375         if (!target__has_cpu(&target) || target__has_per_thread(&target))
 376                 affinity = NULL;
 377         else if (affinity__setup(&saved_affinity) < 0)
 378                 return -1;
 379         else
 380                 affinity = &saved_affinity;
 381
 382         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 383                 struct evsel *counter = evlist_cpu_itr.evsel;
 384
 385                 if (evsel__is_bpf(counter))
 386                         continue;
 387
 388                 if (!counter->err)
 389                         counter->err = read_counter_cpu(counter, evlist_cpu_itr.cpu_map_idx);
 390         }
 391         if (affinity)
 392                 affinity__cleanup(&saved_affinity);
 393
 394         return 0;
 395 }
 396
 397 static int read_bpf_map_counters(void)
 398 {
 399         struct evsel *counter;
 400         int err;
 401
 402         evlist__for_each_entry(evsel_list, counter) {
 403                 if (!evsel__is_bpf(counter))
 404                         continue;
 405
 406                 err = bpf_counter__read(counter);
 407                 if (err)
 408                         return err;
 409         }
 410         return 0;
 411 }
 412
 413 static int read_counters(void)
 414 {
 415         if (!stat_config.stop_read_counter) {
 416                 if (read_bpf_map_counters() ||
 417                     read_affinity_counters())
 418                         return -1;
 419         }
 420         return 0;
 421 }
 422
 423 static void process_counters(void)
 424 {
 425         struct evsel *counter;
 426
 427         evlist__for_each_entry(evsel_list, counter) {
 428                 if (counter->err)
 429                         pr_debug("failed to read counter %s\n", counter->name);
 430                 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
 431                         pr_warning("failed to process counter %s\n", counter->name);
 432                 counter->err = 0;
 433         }
 434
 435         perf_stat_merge_counters(&stat_config, evsel_list);
 436         perf_stat_process_percore(&stat_config, evsel_list);
 437 }
 438
 439 static void process_interval(void)
 440 {
 441         struct timespec ts, rs;
 442
 443         clock_gettime(CLOCK_MONOTONIC, &ts);
 444         diff_timespec(&rs, &ts, &ref_time);
 445
 446         evlist__reset_aggr_stats(evsel_list);
 447
 448         if (read_counters() == 0)
 449                 process_counters();
 450
 451         if (STAT_RECORD) {
 452                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 453                         pr_err("failed to write stat round event\n");
 454         }
 455
 456         init_stats(&walltime_nsecs_stats);
 457         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
 458         print_counters(&rs, 0, NULL);
 459 }
 460
 461 static bool handle_interval(unsigned int interval, int *times)
 462 {
 463         if (interval) {
 464                 process_interval();
 465                 if (interval_count && !(--(*times)))
 466                         return true;
 467         }
 468         return false;
 469 }
 470
 471 static int enable_counters(void)
 472 {
 473         struct evsel *evsel;
 474         int err;
 475
 476         evlist__for_each_entry(evsel_list, evsel) {
 477                 if (!evsel__is_bpf(evsel))
 478                         continue;
 479
 480                 err = bpf_counter__enable(evsel);
 481                 if (err)
 482                         return err;
 483         }
 484
 485         if (!target__enable_on_exec(&target)) {
 486                 if (!all_counters_use_bpf)
 487                         evlist__enable(evsel_list);
 488         }
 489         return 0;
 490 }
 491
 492 static void disable_counters(void)
 493 {
 494         struct evsel *counter;
 495
 496         /*
 497          * If we don't have tracee (attaching to task or cpu), counters may
 498          * still be running. To get accurate group ratios, we must stop groups
 499          * from counting before reading their constituent counters.
 500          */
 501         if (!target__none(&target)) {
 502                 evlist__for_each_entry(evsel_list, counter)
 503                         bpf_counter__disable(counter);
 504                 if (!all_counters_use_bpf)
 505                         evlist__disable(evsel_list);
 506         }
 507 }
 508
 509 static volatile sig_atomic_t workload_exec_errno;
 510
 511 /*
 512  * evlist__prepare_workload will send a SIGUSR1
 513  * if the fork fails, since we asked by setting its
 514  * want_signal to true.
 515  */
 516 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 517                                         void *ucontext __maybe_unused)
 518 {
 519         workload_exec_errno = info->si_value.sival_int;
 520 }
 521
 522 static bool evsel__should_store_id(struct evsel *counter)
 523 {
 524         return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
 525 }
 526
 527 static bool is_target_alive(struct target *_target,
 528                             struct perf_thread_map *threads)
 529 {
 530         struct stat st;
 531         int i;
 532
 533         if (!target__has_task(_target))
 534                 return true;
 535
 536         for (i = 0; i < threads->nr; i++) {
 537                 char path[PATH_MAX];
 538
 539                 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
 540                           threads->map[i].pid);
 541
 542                 if (!stat(path, &st))
 543                         return true;
 544         }
 545
 546         return false;
 547 }
 548
 549 static void process_evlist(struct evlist *evlist, unsigned int interval)
 550 {
 551         enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
 552
 553         if (evlist__ctlfd_process(evlist, &cmd) > 0) {
 554                 switch (cmd) {
 555                 case EVLIST_CTL_CMD_ENABLE:
 556                         fallthrough;
 557                 case EVLIST_CTL_CMD_DISABLE:
 558                         if (interval)
 559                                 process_interval();
 560                         break;
 561                 case EVLIST_CTL_CMD_SNAPSHOT:
 562                 case EVLIST_CTL_CMD_ACK:
 563                 case EVLIST_CTL_CMD_UNSUPPORTED:
 564                 case EVLIST_CTL_CMD_EVLIST:
 565                 case EVLIST_CTL_CMD_STOP:
 566                 case EVLIST_CTL_CMD_PING:
 567                 default:
 568                         break;
 569                 }
 570         }
 571 }
 572
 573 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
 574                         int *time_to_sleep)
 575 {
 576         int tts = *time_to_sleep;
 577         struct timespec time_diff;
 578
 579         diff_timespec(&time_diff, time_stop, time_start);
 580
 581         tts -= time_diff.tv_sec * MSEC_PER_SEC +
 582                time_diff.tv_nsec / NSEC_PER_MSEC;
 583
 584         if (tts < 0)
 585                 tts = 0;
 586
 587         *time_to_sleep = tts;
 588 }
 589
 590 static int dispatch_events(bool forks, int timeout, int interval, int *times)
 591 {
 592         int child_exited = 0, status = 0;
 593         int time_to_sleep, sleep_time;
 594         struct timespec time_start, time_stop;
 595
 596         if (interval)
 597                 sleep_time = interval;
 598         else if (timeout)
 599                 sleep_time = timeout;
 600         else
 601                 sleep_time = 1000;
 602
 603         time_to_sleep = sleep_time;
 604
 605         while (!done) {
 606                 if (forks)
 607                         child_exited = waitpid(child_pid, &status, WNOHANG);
 608                 else
 609                         child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
 610
 611                 if (child_exited)
 612                         break;
 613
 614                 clock_gettime(CLOCK_MONOTONIC, &time_start);
 615                 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
 616                         if (timeout || handle_interval(interval, times))
 617                                 break;
 618                         time_to_sleep = sleep_time;
 619                 } else { /* fd revent */
 620                         process_evlist(evsel_list, interval);
 621                         clock_gettime(CLOCK_MONOTONIC, &time_stop);
 622                         compute_tts(&time_start, &time_stop, &time_to_sleep);
 623                 }
 624         }
 625
 626         return status;
 627 }
 628
 629 enum counter_recovery {
 630         COUNTER_SKIP,
 631         COUNTER_RETRY,
 632         COUNTER_FATAL,
 633 };
 634
 635 static enum counter_recovery stat_handle_error(struct evsel *counter)
 636 {
 637         char msg[BUFSIZ];
 638         /*
 639          * PPC returns ENXIO for HW counters until 2.6.37
 640          * (behavior changed with commit b0a873e).
 641          */
 642         if (errno == EINVAL || errno == ENOSYS ||
 643             errno == ENOENT || errno == ENXIO) {
 644                 if (verbose > 0)
 645                         ui__warning("%s event is not supported by the kernel.\n",
 646                                     evsel__name(counter));
 647                 counter->supported = false;
 648                 /*
 649                  * errored is a sticky flag that means one of the counter's
 650                  * cpu event had a problem and needs to be reexamined.
 651                  */
 652                 counter->errored = true;
 653
 654                 if ((evsel__leader(counter) != counter) ||
 655                     !(counter->core.leader->nr_members > 1))
 656                         return COUNTER_SKIP;
 657         } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
 658                 if (verbose > 0)
 659                         ui__warning("%s\n", msg);
 660                 return COUNTER_RETRY;
 661         } else if (target__has_per_thread(&target) && errno != EOPNOTSUPP &&
 662                    evsel_list->core.threads &&
 663                    evsel_list->core.threads->err_thread != -1) {
 664                 /*
 665                  * For global --per-thread case, skip current
 666                  * error thread.
 667                  */
 668                 if (!thread_map__remove(evsel_list->core.threads,
 669                                         evsel_list->core.threads->err_thread)) {
 670                         evsel_list->core.threads->err_thread = -1;
 671                         return COUNTER_RETRY;
 672                 }
 673         } else if (counter->skippable) {
 674                 if (verbose > 0)
 675                         ui__warning("skipping event %s that kernel failed to open .\n",
 676                                     evsel__name(counter));
 677                 counter->supported = false;
 678                 counter->errored = true;
 679                 return COUNTER_SKIP;
 680         }
 681
 682         if (errno == EOPNOTSUPP) {
 683                 if (verbose > 0) {
 684                         ui__warning("%s event is not supported by the kernel.\n",
 685                                     evsel__name(counter));
 686                 }
 687                 counter->supported = false;
 688                 counter->errored = true;
 689
 690                 if ((evsel__leader(counter) != counter) ||
 691                     !(counter->core.leader->nr_members > 1))
 692                         return COUNTER_SKIP;
 693         }
 694
 695         evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
 696         ui__error("%s\n", msg);
 697
 698         if (child_pid != -1)
 699                 kill(child_pid, SIGTERM);
 700
 701         tpebs_delete();
 702
 703         return COUNTER_FATAL;
 704 }
 705
 706 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 707 {
 708         int interval = stat_config.interval;
 709         int times = stat_config.times;
 710         int timeout = stat_config.timeout;
 711         char msg[BUFSIZ];
 712         unsigned long long t0, t1;
 713         struct evsel *counter;
 714         size_t l;
 715         int status = 0;
 716         const bool forks = (argc > 0);
 717         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 718         struct evlist_cpu_iterator evlist_cpu_itr;
 719         struct affinity saved_affinity, *affinity = NULL;
 720         int err;
 721         bool second_pass = false;
 722
 723         if (forks) {
 724                 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
 725                         perror("failed to prepare workload");
 726                         return -1;
 727                 }
 728                 child_pid = evsel_list->workload.pid;
 729         }
 730
 731         if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
 732                 if (affinity__setup(&saved_affinity) < 0) {
 733                         err = -1;
 734                         goto err_out;
 735                 }
 736                 affinity = &saved_affinity;
 737         }
 738
 739         evlist__for_each_entry(evsel_list, counter) {
 740                 counter->reset_group = false;
 741                 if (bpf_counter__load(counter, &target)) {
 742                         err = -1;
 743                         goto err_out;
 744                 }
 745                 if (!(evsel__is_bperf(counter)))
 746                         all_counters_use_bpf = false;
 747         }
 748
 749         evlist__reset_aggr_stats(evsel_list);
 750
 751         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 752                 counter = evlist_cpu_itr.evsel;
 753
 754                 /*
 755                  * bperf calls evsel__open_per_cpu() in bperf__load(), so
 756                  * no need to call it again here.
 757                  */
 758                 if (target.use_bpf)
 759                         break;
 760
 761                 if (counter->reset_group || counter->errored)
 762                         continue;
 763                 if (evsel__is_bperf(counter))
 764                         continue;
 765 try_again:
 766                 if (create_perf_stat_counter(counter, &stat_config, &target,
 767                                              evlist_cpu_itr.cpu_map_idx) < 0) {
 768
 769                         /*
 770                          * Weak group failed. We cannot just undo this here
 771                          * because earlier CPUs might be in group mode, and the kernel
 772                          * doesn't support mixing group and non group reads. Defer
 773                          * it to later.
 774                          * Don't close here because we're in the wrong affinity.
 775                          */
 776                         if ((errno == EINVAL || errno == EBADF) &&
 777                                 evsel__leader(counter) != counter &&
 778                                 counter->weak_group) {
 779                                 evlist__reset_weak_group(evsel_list, counter, false);
 780                                 assert(counter->reset_group);
 781                                 second_pass = true;
 782                                 continue;
 783                         }
 784
 785                         switch (stat_handle_error(counter)) {
 786                         case COUNTER_FATAL:
 787                                 err = -1;
 788                                 goto err_out;
 789                         case COUNTER_RETRY:
 790                                 goto try_again;
 791                         case COUNTER_SKIP:
 792                                 continue;
 793                         default:
 794                                 break;
 795                         }
 796
 797                 }
 798                 counter->supported = true;
 799         }
 800
 801         if (second_pass) {
 802                 /*
 803                  * Now redo all the weak group after closing them,
 804                  * and also close errored counters.
 805                  */
 806
 807                 /* First close errored or weak retry */
 808                 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 809                         counter = evlist_cpu_itr.evsel;
 810
 811                         if (!counter->reset_group && !counter->errored)
 812                                 continue;
 813
 814                         perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
 815                 }
 816                 /* Now reopen weak */
 817                 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 818                         counter = evlist_cpu_itr.evsel;
 819
 820                         if (!counter->reset_group)
 821                                 continue;
 822 try_again_reset:
 823                         pr_debug2("reopening weak %s\n", evsel__name(counter));
 824                         if (create_perf_stat_counter(counter, &stat_config, &target,
 825                                                      evlist_cpu_itr.cpu_map_idx) < 0) {
 826
 827                                 switch (stat_handle_error(counter)) {
 828                                 case COUNTER_FATAL:
 829                                         err = -1;
 830                                         goto err_out;
 831                                 case COUNTER_RETRY:
 832                                         goto try_again_reset;
 833                                 case COUNTER_SKIP:
 834                                         continue;
 835                                 default:
 836                                         break;
 837                                 }
 838                         }
 839                         counter->supported = true;
 840                 }
 841         }
 842         affinity__cleanup(affinity);
 843         affinity = NULL;
 844
 845         evlist__for_each_entry(evsel_list, counter) {
 846                 if (!counter->supported) {
 847                         perf_evsel__free_fd(&counter->core);
 848                         continue;
 849                 }
 850
 851                 l = strlen(counter->unit);
 852                 if (l > stat_config.unit_width)
 853                         stat_config.unit_width = l;
 854
 855                 if (evsel__should_store_id(counter) &&
 856                     evsel__store_ids(counter, evsel_list)) {
 857                         err = -1;
 858                         goto err_out;
 859                 }
 860         }
 861
 862         if (evlist__apply_filters(evsel_list, &counter, &target)) {
 863                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 864                         counter->filter, evsel__name(counter), errno,
 865                         str_error_r(errno, msg, sizeof(msg)));
 866                 return -1;
 867         }
 868
 869         if (STAT_RECORD) {
 870                 int fd = perf_data__fd(&perf_stat.data);
 871
 872                 if (is_pipe) {
 873                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 874                 } else {
 875                         err = perf_session__write_header(perf_stat.session, evsel_list,
 876                                                          fd, false);
 877                 }
 878
 879                 if (err < 0)
 880                         goto err_out;
 881
 882                 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
 883                                                          process_synthesized_event, is_pipe);
 884                 if (err < 0)
 885                         goto err_out;
 886
 887         }
 888
 889         if (target.initial_delay) {
 890                 pr_info(EVLIST_DISABLED_MSG);
 891         } else {
 892                 err = enable_counters();
 893                 if (err) {
 894                         err = -1;
 895                         goto err_out;
 896                 }
 897         }
 898
 899         /* Exec the command, if any */
 900         if (forks)
 901                 evlist__start_workload(evsel_list);
 902
 903         if (target.initial_delay > 0) {
 904                 usleep(target.initial_delay * USEC_PER_MSEC);
 905                 err = enable_counters();
 906                 if (err) {
 907                         err = -1;
 908                         goto err_out;
 909                 }
 910
 911                 pr_info(EVLIST_ENABLED_MSG);
 912         }
 913
 914         t0 = rdclock();
 915         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 916
 917         if (forks) {
 918                 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
 919                         status = dispatch_events(forks, timeout, interval, &times);
 920                 if (child_pid != -1) {
 921                         if (timeout)
 922                                 kill(child_pid, SIGTERM);
 923                         wait4(child_pid, &status, 0, &stat_config.ru_data);
 924                 }
 925
 926                 if (workload_exec_errno) {
 927                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 928                         pr_err("Workload failed: %s\n", emsg);
 929                         err = -1;
 930                         goto err_out;
 931                 }
 932
 933                 if (WIFSIGNALED(status))
 934                         psignal(WTERMSIG(status), argv[0]);
 935         } else {
 936                 status = dispatch_events(forks, timeout, interval, &times);
 937         }
 938
 939         disable_counters();
 940
 941         t1 = rdclock();
 942
 943         if (stat_config.walltime_run_table)
 944                 stat_config.walltime_run[run_idx] = t1 - t0;
 945
 946         if (interval && stat_config.summary) {
 947                 stat_config.interval = 0;
 948                 stat_config.stop_read_counter = true;
 949                 init_stats(&walltime_nsecs_stats);
 950                 update_stats(&walltime_nsecs_stats, t1 - t0);
 951
 952                 evlist__copy_prev_raw_counts(evsel_list);
 953                 evlist__reset_prev_raw_counts(evsel_list);
 954                 evlist__reset_aggr_stats(evsel_list);
 955         } else {
 956                 update_stats(&walltime_nsecs_stats, t1 - t0);
 957                 update_rusage_stats(&ru_stats, &stat_config.ru_data);
 958         }
 959
 960         /*
 961          * Closing a group leader splits the group, and as we only disable
 962          * group leaders, results in remaining events becoming enabled. To
 963          * avoid arbitrary skew, we must read all counters before closing any
 964          * group leaders.
 965          */
 966         if (read_counters() == 0)
 967                 process_counters();
 968
 969         /*
 970          * We need to keep evsel_list alive, because it's processed
 971          * later the evsel_list will be closed after.
 972          */
 973         if (!STAT_RECORD)
 974                 evlist__close(evsel_list);
 975
 976         return WEXITSTATUS(status);
 977
 978 err_out:
 979         if (forks)
 980                 evlist__cancel_workload(evsel_list);
 981
 982         affinity__cleanup(affinity);
 983         return err;
 984 }
 985
 986 /*
 987  * Returns -1 for fatal errors which signifies to not continue
 988  * when in repeat mode.
 989  *
 990  * Returns < -1 error codes when stat record is used. These
 991  * result in the stat information being displayed, but writing
 992  * to the file fails and is non fatal.
 993  */
 994 static int run_perf_stat(int argc, const char **argv, int run_idx)
 995 {
 996         int ret;
 997
 998         if (pre_cmd) {
 999                 ret = system(pre_cmd);
1000                 if (ret)
1001                         return ret;
1002         }
1003
1004         if (sync_run)
1005                 sync();
1006
1007         ret = __run_perf_stat(argc, argv, run_idx);
1008         if (ret)
1009                 return ret;
1010
1011         if (post_cmd) {
1012                 ret = system(post_cmd);
1013                 if (ret)
1014                         return ret;
1015         }
1016
1017         return ret;
1018 }
1019
1020 static void print_counters(struct timespec *ts, int argc, const char **argv)
1021 {
1022         /* Do not print anything if we record to the pipe. */
1023         if (STAT_RECORD && perf_stat.data.is_pipe)
1024                 return;
1025         if (quiet)
1026                 return;
1027
1028         evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1029 }
1030
1031 static volatile sig_atomic_t signr = -1;
1032
1033 static void skip_signal(int signo)
1034 {
1035         if ((child_pid == -1) || stat_config.interval)
1036                 done = 1;
1037
1038         signr = signo;
1039         /*
1040          * render child_pid harmless
1041          * won't send SIGTERM to a random
1042          * process in case of race condition
1043          * and fast PID recycling
1044          */
1045         child_pid = -1;
1046 }
1047
1048 static void sig_atexit(void)
1049 {
1050         sigset_t set, oset;
1051
1052         /*
1053          * avoid race condition with SIGCHLD handler
1054          * in skip_signal() which is modifying child_pid
1055          * goal is to avoid send SIGTERM to a random
1056          * process
1057          */
1058         sigemptyset(&set);
1059         sigaddset(&set, SIGCHLD);
1060         sigprocmask(SIG_BLOCK, &set, &oset);
1061
1062         if (child_pid != -1)
1063                 kill(child_pid, SIGTERM);
1064
1065         sigprocmask(SIG_SETMASK, &oset, NULL);
1066
1067         if (signr == -1)
1068                 return;
1069
1070         signal(signr, SIG_DFL);
1071         kill(getpid(), signr);
1072 }
1073
1074 void perf_stat__set_big_num(int set)
1075 {
1076         stat_config.big_num = (set != 0);
1077 }
1078
1079 void perf_stat__set_no_csv_summary(int set)
1080 {
1081         stat_config.no_csv_summary = (set != 0);
1082 }
1083
1084 static int stat__set_big_num(const struct option *opt __maybe_unused,
1085                              const char *s __maybe_unused, int unset)
1086 {
1087         big_num_opt = unset ? 0 : 1;
1088         perf_stat__set_big_num(!unset);
1089         return 0;
1090 }
1091
1092 static int enable_metric_only(const struct option *opt __maybe_unused,
1093                               const char *s __maybe_unused, int unset)
1094 {
1095         force_metric_only = true;
1096         stat_config.metric_only = !unset;
1097         return 0;
1098 }
1099
1100 static int append_metric_groups(const struct option *opt __maybe_unused,
1101                                const char *str,
1102                                int unset __maybe_unused)
1103 {
1104         if (metrics) {
1105                 char *tmp;
1106
1107                 if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1108                         return -ENOMEM;
1109                 free(metrics);
1110                 metrics = tmp;
1111         } else {
1112                 metrics = strdup(str);
1113                 if (!metrics)
1114                         return -ENOMEM;
1115         }
1116         return 0;
1117 }
1118
1119 static int parse_control_option(const struct option *opt,
1120                                 const char *str,
1121                                 int unset __maybe_unused)
1122 {
1123         struct perf_stat_config *config = opt->value;
1124
1125         return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1126 }
1127
1128 static int parse_stat_cgroups(const struct option *opt,
1129                               const char *str, int unset)
1130 {
1131         if (stat_config.cgroup_list) {
1132                 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1133                 return -1;
1134         }
1135
1136         return parse_cgroups(opt, str, unset);
1137 }
1138
1139 static int parse_cputype(const struct option *opt,
1140                              const char *str,
1141                              int unset __maybe_unused)
1142 {
1143         const struct perf_pmu *pmu;
1144         struct evlist *evlist = *(struct evlist **)opt->value;
1145
1146         if (!list_empty(&evlist->core.entries)) {
1147                 fprintf(stderr, "Must define cputype before events/metrics\n");
1148                 return -1;
1149         }
1150
1151         pmu = perf_pmus__pmu_for_pmu_filter(str);
1152         if (!pmu) {
1153                 fprintf(stderr, "--cputype %s is not supported!\n", str);
1154                 return -1;
1155         }
1156         parse_events_option_args.pmu_filter = pmu->name;
1157
1158         return 0;
1159 }
1160
1161 static int parse_cache_level(const struct option *opt,
1162                              const char *str,
1163                              int unset __maybe_unused)
1164 {
1165         int level;
1166         struct opt_aggr_mode *opt_aggr_mode = (struct opt_aggr_mode *)opt->value;
1167         u32 *aggr_level = (u32 *)opt->data;
1168
1169         /*
1170          * If no string is specified, aggregate based on the topology of
1171          * Last Level Cache (LLC). Since the LLC level can change from
1172          * architecture to architecture, set level greater than
1173          * MAX_CACHE_LVL which will be interpreted as LLC.
1174          */
1175         if (str == NULL) {
1176                 level = MAX_CACHE_LVL + 1;
1177                 goto out;
1178         }
1179
1180         /*
1181          * The format to specify cache level is LX or lX where X is the
1182          * cache level.
1183          */
1184         if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
1185                 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1186                        MAX_CACHE_LVL,
1187                        MAX_CACHE_LVL);
1188                 return -EINVAL;
1189         }
1190
1191         level = atoi(&str[1]);
1192         if (level < 1) {
1193                 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1194                        MAX_CACHE_LVL,
1195                        MAX_CACHE_LVL);
1196                 return -EINVAL;
1197         }
1198
1199         if (level > MAX_CACHE_LVL) {
1200                 pr_err("perf only supports max cache level of %d.\n"
1201                        "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
1202                 return -EINVAL;
1203         }
1204 out:
1205         opt_aggr_mode->cache = true;
1206         *aggr_level = level;
1207         return 0;
1208 }
1209
1210 /**
1211  * Calculate the cache instance ID from the map in
1212  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1213  * Cache instance ID is the first CPU reported in the shared_cpu_list file.
1214  */
1215 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
1216 {
1217         int id;
1218         struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
1219
1220         /*
1221          * If the map contains no CPU, consider the current CPU to
1222          * be the first online CPU in the cache domain else use the
1223          * first online CPU of the cache domain as the ID.
1224          */
1225         id = perf_cpu_map__min(cpu_map).cpu;
1226         if (id == -1)
1227                 id = cpu.cpu;
1228
1229         /* Free the perf_cpu_map used to find the cache ID */
1230         perf_cpu_map__put(cpu_map);
1231
1232         return id;
1233 }
1234
1235 /**
1236  * cpu__get_cache_id - Returns 0 if successful in populating the
1237  * cache level and cache id. Cache level is read from
1238  * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
1239  * is the first CPU reported by
1240  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1241  */
1242 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
1243 {
1244         int ret = 0;
1245         u32 cache_level = stat_config.aggr_level;
1246         struct cpu_cache_level caches[MAX_CACHE_LVL];
1247         u32 i = 0, caches_cnt = 0;
1248
1249         cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1250         cache->cache = -1;
1251
1252         ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
1253         if (ret) {
1254                 /*
1255                  * If caches_cnt is not 0, cpu_cache_level data
1256                  * was allocated when building the topology.
1257                  * Free the allocated data before returning.
1258                  */
1259                 if (caches_cnt)
1260                         goto free_caches;
1261
1262                 return ret;
1263         }
1264
1265         if (!caches_cnt)
1266                 return -1;
1267
1268         /*
1269          * Save the data for the highest level if no
1270          * level was specified by the user.
1271          */
1272         if (cache_level > MAX_CACHE_LVL) {
1273                 int max_level_index = 0;
1274
1275                 for (i = 1; i < caches_cnt; ++i) {
1276                         if (caches[i].level > caches[max_level_index].level)
1277                                 max_level_index = i;
1278                 }
1279
1280                 cache->cache_lvl = caches[max_level_index].level;
1281                 cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
1282
1283                 /* Reset i to 0 to free entire caches[] */
1284                 i = 0;
1285                 goto free_caches;
1286         }
1287
1288         for (i = 0; i < caches_cnt; ++i) {
1289                 if (caches[i].level == cache_level) {
1290                         cache->cache_lvl = cache_level;
1291                         cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1292                 }
1293
1294                 cpu_cache_level__free(&caches[i]);
1295         }
1296
1297 free_caches:
1298         /*
1299          * Free all the allocated cpu_cache_level data.
1300          */
1301         while (i < caches_cnt)
1302                 cpu_cache_level__free(&caches[i++]);
1303
1304         return ret;
1305 }
1306
1307 /**
1308  * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
1309  * level, die and socket populated with the cache instache ID, cache level,
1310  * die and socket for cpu. The function signature is compatible with
1311  * aggr_cpu_id_get_t.
1312  */
1313 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
1314 {
1315         int ret;
1316         struct aggr_cpu_id id;
1317         struct perf_cache cache;
1318
1319         id = aggr_cpu_id__die(cpu, data);
1320         if (aggr_cpu_id__is_empty(&id))
1321                 return id;
1322
1323         ret = cpu__get_cache_details(cpu, &cache);
1324         if (ret)
1325                 return id;
1326
1327         id.cache_lvl = cache.cache_lvl;
1328         id.cache = cache.cache;
1329         return id;
1330 }
1331
1332 static const char *const aggr_mode__string[] = {
1333         [AGGR_CORE] = "core",
1334         [AGGR_CACHE] = "cache",
1335         [AGGR_CLUSTER] = "cluster",
1336         [AGGR_DIE] = "die",
1337         [AGGR_GLOBAL] = "global",
1338         [AGGR_NODE] = "node",
1339         [AGGR_NONE] = "none",
1340         [AGGR_SOCKET] = "socket",
1341         [AGGR_THREAD] = "thread",
1342         [AGGR_UNSET] = "unset",
1343 };
1344
1345 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1346                                                 struct perf_cpu cpu)
1347 {
1348         return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1349 }
1350
1351 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1352                                              struct perf_cpu cpu)
1353 {
1354         return aggr_cpu_id__die(cpu, /*data=*/NULL);
1355 }
1356
1357 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
1358                                                   struct perf_cpu cpu)
1359 {
1360         return aggr_cpu_id__cache(cpu, /*data=*/NULL);
1361 }
1362
1363 static struct aggr_cpu_id perf_stat__get_cluster(struct perf_stat_config *config __maybe_unused,
1364                                                  struct perf_cpu cpu)
1365 {
1366         return aggr_cpu_id__cluster(cpu, /*data=*/NULL);
1367 }
1368
1369 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1370                                               struct perf_cpu cpu)
1371 {
1372         return aggr_cpu_id__core(cpu, /*data=*/NULL);
1373 }
1374
1375 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1376                                               struct perf_cpu cpu)
1377 {
1378         return aggr_cpu_id__node(cpu, /*data=*/NULL);
1379 }
1380
1381 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1382                                                 struct perf_cpu cpu)
1383 {
1384         return aggr_cpu_id__global(cpu, /*data=*/NULL);
1385 }
1386
1387 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1388                                              struct perf_cpu cpu)
1389 {
1390         return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1391 }
1392
1393 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1394                                               aggr_get_id_t get_id, struct perf_cpu cpu)
1395 {
1396         struct aggr_cpu_id id;
1397
1398         /* per-process mode - should use global aggr mode */
1399         if (cpu.cpu == -1)
1400                 return get_id(config, cpu);
1401
1402         if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1403                 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1404
1405         id = config->cpus_aggr_map->map[cpu.cpu];
1406         return id;
1407 }
1408
1409 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1410                                                        struct perf_cpu cpu)
1411 {
1412         return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1413 }
1414
1415 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1416                                                     struct perf_cpu cpu)
1417 {
1418         return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1419 }
1420
1421 static struct aggr_cpu_id perf_stat__get_cluster_cached(struct perf_stat_config *config,
1422                                                         struct perf_cpu cpu)
1423 {
1424         return perf_stat__get_aggr(config, perf_stat__get_cluster, cpu);
1425 }
1426
1427 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
1428                                                          struct perf_cpu cpu)
1429 {
1430         return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
1431 }
1432
1433 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1434                                                      struct perf_cpu cpu)
1435 {
1436         return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1437 }
1438
1439 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1440                                                      struct perf_cpu cpu)
1441 {
1442         return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1443 }
1444
1445 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1446                                                        struct perf_cpu cpu)
1447 {
1448         return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1449 }
1450
1451 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1452                                                     struct perf_cpu cpu)
1453 {
1454         return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1455 }
1456
1457 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1458 {
1459         switch (aggr_mode) {
1460         case AGGR_SOCKET:
1461                 return aggr_cpu_id__socket;
1462         case AGGR_DIE:
1463                 return aggr_cpu_id__die;
1464         case AGGR_CLUSTER:
1465                 return aggr_cpu_id__cluster;
1466         case AGGR_CACHE:
1467                 return aggr_cpu_id__cache;
1468         case AGGR_CORE:
1469                 return aggr_cpu_id__core;
1470         case AGGR_NODE:
1471                 return aggr_cpu_id__node;
1472         case AGGR_NONE:
1473                 return aggr_cpu_id__cpu;
1474         case AGGR_GLOBAL:
1475                 return aggr_cpu_id__global;
1476         case AGGR_THREAD:
1477         case AGGR_UNSET:
1478         case AGGR_MAX:
1479         default:
1480                 return NULL;
1481         }
1482 }
1483
1484 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1485 {
1486         switch (aggr_mode) {
1487         case AGGR_SOCKET:
1488                 return perf_stat__get_socket_cached;
1489         case AGGR_DIE:
1490                 return perf_stat__get_die_cached;
1491         case AGGR_CLUSTER:
1492                 return perf_stat__get_cluster_cached;
1493         case AGGR_CACHE:
1494                 return perf_stat__get_cache_id_cached;
1495         case AGGR_CORE:
1496                 return perf_stat__get_core_cached;
1497         case AGGR_NODE:
1498                 return perf_stat__get_node_cached;
1499         case AGGR_NONE:
1500                 return perf_stat__get_cpu_cached;
1501         case AGGR_GLOBAL:
1502                 return perf_stat__get_global_cached;
1503         case AGGR_THREAD:
1504         case AGGR_UNSET:
1505         case AGGR_MAX:
1506         default:
1507                 return NULL;
1508         }
1509 }
1510
1511 static int perf_stat_init_aggr_mode(void)
1512 {
1513         int nr;
1514         aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1515
1516         if (get_id) {
1517                 bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1518                 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1519                                                          get_id, /*data=*/NULL, needs_sort);
1520                 if (!stat_config.aggr_map) {
1521                         pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1522                         return -1;
1523                 }
1524                 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1525         }
1526
1527         if (stat_config.aggr_mode == AGGR_THREAD) {
1528                 nr = perf_thread_map__nr(evsel_list->core.threads);
1529                 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1530                 if (stat_config.aggr_map == NULL)
1531                         return -ENOMEM;
1532
1533                 for (int s = 0; s < nr; s++) {
1534                         struct aggr_cpu_id id = aggr_cpu_id__empty();
1535
1536                         id.thread_idx = s;
1537                         stat_config.aggr_map->map[s] = id;
1538                 }
1539                 return 0;
1540         }
1541
1542         /*
1543          * The evsel_list->cpus is the base we operate on,
1544          * taking the highest cpu number to be the size of
1545          * the aggregation translate cpumap.
1546          */
1547         if (!perf_cpu_map__is_any_cpu_or_is_empty(evsel_list->core.user_requested_cpus))
1548                 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1549         else
1550                 nr = 0;
1551         stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1552         return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1553 }
1554
1555 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1556 {
1557         free(map);
1558 }
1559
1560 static void perf_stat__exit_aggr_mode(void)
1561 {
1562         cpu_aggr_map__delete(stat_config.aggr_map);
1563         cpu_aggr_map__delete(stat_config.cpus_aggr_map);
1564         stat_config.aggr_map = NULL;
1565         stat_config.cpus_aggr_map = NULL;
1566 }
1567
1568 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1569 {
1570         struct perf_env *env = data;
1571         struct aggr_cpu_id id = aggr_cpu_id__empty();
1572
1573         if (cpu.cpu != -1)
1574                 id.socket = env->cpu[cpu.cpu].socket_id;
1575
1576         return id;
1577 }
1578
1579 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1580 {
1581         struct perf_env *env = data;
1582         struct aggr_cpu_id id = aggr_cpu_id__empty();
1583
1584         if (cpu.cpu != -1) {
1585                 /*
1586                  * die_id is relative to socket, so start
1587                  * with the socket ID and then add die to
1588                  * make a unique ID.
1589                  */
1590                 id.socket = env->cpu[cpu.cpu].socket_id;
1591                 id.die = env->cpu[cpu.cpu].die_id;
1592         }
1593
1594         return id;
1595 }
1596
1597 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
1598                                            u32 cache_level, struct aggr_cpu_id *id)
1599 {
1600         int i;
1601         int caches_cnt = env->caches_cnt;
1602         struct cpu_cache_level *caches = env->caches;
1603
1604         id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1605         id->cache = -1;
1606
1607         if (!caches_cnt)
1608                 return;
1609
1610         for (i = caches_cnt - 1; i > -1; --i) {
1611                 struct perf_cpu_map *cpu_map;
1612                 int map_contains_cpu;
1613
1614                 /*
1615                  * If user has not specified a level, find the fist level with
1616                  * the cpu in the map. Since building the map is expensive, do
1617                  * this only if levels match.
1618                  */
1619                 if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
1620                         continue;
1621
1622                 cpu_map = perf_cpu_map__new(caches[i].map);
1623                 map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
1624                 perf_cpu_map__put(cpu_map);
1625
1626                 if (map_contains_cpu != -1) {
1627                         id->cache_lvl = caches[i].level;
1628                         id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1629                         return;
1630                 }
1631         }
1632 }
1633
1634 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
1635                                                           void *data)
1636 {
1637         struct perf_env *env = data;
1638         struct aggr_cpu_id id = aggr_cpu_id__empty();
1639
1640         if (cpu.cpu != -1) {
1641                 u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
1642
1643                 id.socket = env->cpu[cpu.cpu].socket_id;
1644                 id.die = env->cpu[cpu.cpu].die_id;
1645                 perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
1646         }
1647
1648         return id;
1649 }
1650
1651 static struct aggr_cpu_id perf_env__get_cluster_aggr_by_cpu(struct perf_cpu cpu,
1652                                                             void *data)
1653 {
1654         struct perf_env *env = data;
1655         struct aggr_cpu_id id = aggr_cpu_id__empty();
1656
1657         if (cpu.cpu != -1) {
1658                 id.socket = env->cpu[cpu.cpu].socket_id;
1659                 id.die = env->cpu[cpu.cpu].die_id;
1660                 id.cluster = env->cpu[cpu.cpu].cluster_id;
1661         }
1662
1663         return id;
1664 }
1665
1666 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1667 {
1668         struct perf_env *env = data;
1669         struct aggr_cpu_id id = aggr_cpu_id__empty();
1670
1671         if (cpu.cpu != -1) {
1672                 /*
1673                  * core_id is relative to socket, die and cluster, we need a
1674                  * global id. So we set socket, die id, cluster id and core id.
1675                  */
1676                 id.socket = env->cpu[cpu.cpu].socket_id;
1677                 id.die = env->cpu[cpu.cpu].die_id;
1678                 id.cluster = env->cpu[cpu.cpu].cluster_id;
1679                 id.core = env->cpu[cpu.cpu].core_id;
1680         }
1681
1682         return id;
1683 }
1684
1685 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1686 {
1687         struct perf_env *env = data;
1688         struct aggr_cpu_id id = aggr_cpu_id__empty();
1689
1690         if (cpu.cpu != -1) {
1691                 /*
1692                  * core_id is relative to socket and die,
1693                  * we need a global id. So we set
1694                  * socket, die id and core id
1695                  */
1696                 id.socket = env->cpu[cpu.cpu].socket_id;
1697                 id.die = env->cpu[cpu.cpu].die_id;
1698                 id.core = env->cpu[cpu.cpu].core_id;
1699                 id.cpu = cpu;
1700         }
1701
1702         return id;
1703 }
1704
1705 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1706 {
1707         struct aggr_cpu_id id = aggr_cpu_id__empty();
1708
1709         id.node = perf_env__numa_node(data, cpu);
1710         return id;
1711 }
1712
1713 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1714                                                            void *data __maybe_unused)
1715 {
1716         struct aggr_cpu_id id = aggr_cpu_id__empty();
1717
1718         /* it always aggregates to the cpu 0 */
1719         id.cpu = (struct perf_cpu){ .cpu = 0 };
1720         return id;
1721 }
1722
1723 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1724                                                      struct perf_cpu cpu)
1725 {
1726         return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1727 }
1728 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1729                                                   struct perf_cpu cpu)
1730 {
1731         return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1732 }
1733
1734 static struct aggr_cpu_id perf_stat__get_cluster_file(struct perf_stat_config *config __maybe_unused,
1735                                                       struct perf_cpu cpu)
1736 {
1737         return perf_env__get_cluster_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1738 }
1739
1740 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
1741                                                     struct perf_cpu cpu)
1742 {
1743         return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1744 }
1745
1746 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1747                                                    struct perf_cpu cpu)
1748 {
1749         return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1750 }
1751
1752 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1753                                                   struct perf_cpu cpu)
1754 {
1755         return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1756 }
1757
1758 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1759                                                    struct perf_cpu cpu)
1760 {
1761         return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1762 }
1763
1764 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1765                                                      struct perf_cpu cpu)
1766 {
1767         return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1768 }
1769
1770 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1771 {
1772         switch (aggr_mode) {
1773         case AGGR_SOCKET:
1774                 return perf_env__get_socket_aggr_by_cpu;
1775         case AGGR_DIE:
1776                 return perf_env__get_die_aggr_by_cpu;
1777         case AGGR_CLUSTER:
1778                 return perf_env__get_cluster_aggr_by_cpu;
1779         case AGGR_CACHE:
1780                 return perf_env__get_cache_aggr_by_cpu;
1781         case AGGR_CORE:
1782                 return perf_env__get_core_aggr_by_cpu;
1783         case AGGR_NODE:
1784                 return perf_env__get_node_aggr_by_cpu;
1785         case AGGR_GLOBAL:
1786                 return perf_env__get_global_aggr_by_cpu;
1787         case AGGR_NONE:
1788                 return perf_env__get_cpu_aggr_by_cpu;
1789         case AGGR_THREAD:
1790         case AGGR_UNSET:
1791         case AGGR_MAX:
1792         default:
1793                 return NULL;
1794         }
1795 }
1796
1797 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1798 {
1799         switch (aggr_mode) {
1800         case AGGR_SOCKET:
1801                 return perf_stat__get_socket_file;
1802         case AGGR_DIE:
1803                 return perf_stat__get_die_file;
1804         case AGGR_CLUSTER:
1805                 return perf_stat__get_cluster_file;
1806         case AGGR_CACHE:
1807                 return perf_stat__get_cache_file;
1808         case AGGR_CORE:
1809                 return perf_stat__get_core_file;
1810         case AGGR_NODE:
1811                 return perf_stat__get_node_file;
1812         case AGGR_GLOBAL:
1813                 return perf_stat__get_global_file;
1814         case AGGR_NONE:
1815                 return perf_stat__get_cpu_file;
1816         case AGGR_THREAD:
1817         case AGGR_UNSET:
1818         case AGGR_MAX:
1819         default:
1820                 return NULL;
1821         }
1822 }
1823
1824 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1825 {
1826         struct perf_env *env = &st->session->header.env;
1827         aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1828         bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1829
1830         if (stat_config.aggr_mode == AGGR_THREAD) {
1831                 int nr = perf_thread_map__nr(evsel_list->core.threads);
1832
1833                 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1834                 if (stat_config.aggr_map == NULL)
1835                         return -ENOMEM;
1836
1837                 for (int s = 0; s < nr; s++) {
1838                         struct aggr_cpu_id id = aggr_cpu_id__empty();
1839
1840                         id.thread_idx = s;
1841                         stat_config.aggr_map->map[s] = id;
1842                 }
1843                 return 0;
1844         }
1845
1846         if (!get_id)
1847                 return 0;
1848
1849         stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1850                                                  get_id, env, needs_sort);
1851         if (!stat_config.aggr_map) {
1852                 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1853                 return -1;
1854         }
1855         stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1856         return 0;
1857 }
1858
1859 /*
1860  * Add default events, if there were no attributes specified or
1861  * if -d/--detailed, -d -d or -d -d -d is used:
1862  */
1863 static int add_default_events(void)
1864 {
1865         const char *pmu = parse_events_option_args.pmu_filter ?: "all";
1866         struct parse_events_error err;
1867         struct evlist *evlist = evlist__new();
1868         struct evsel *evsel;
1869         int ret = 0;
1870
1871         if (!evlist)
1872                 return -ENOMEM;
1873
1874         parse_events_error__init(&err);
1875
1876         /* Set attrs if no event is selected and !null_run: */
1877         if (stat_config.null_run)
1878                 goto out;
1879
1880         if (transaction_run) {
1881                 /* Handle -T as -M transaction. Once platform specific metrics
1882                  * support has been added to the json files, all architectures
1883                  * will use this approach. To determine transaction support
1884                  * on an architecture test for such a metric name.
1885                  */
1886                 if (!metricgroup__has_metric(pmu, "transaction")) {
1887                         pr_err("Missing transaction metrics\n");
1888                         ret = -1;
1889                         goto out;
1890                 }
1891                 ret = metricgroup__parse_groups(evlist, pmu, "transaction",
1892                                                 stat_config.metric_no_group,
1893                                                 stat_config.metric_no_merge,
1894                                                 stat_config.metric_no_threshold,
1895                                                 stat_config.user_requested_cpu_list,
1896                                                 stat_config.system_wide,
1897                                                 stat_config.hardware_aware_grouping,
1898                                                 &stat_config.metric_events);
1899                 goto out;
1900         }
1901
1902         if (smi_cost) {
1903                 int smi;
1904
1905                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1906                         pr_err("freeze_on_smi is not supported.\n");
1907                         ret = -1;
1908                         goto out;
1909                 }
1910
1911                 if (!smi) {
1912                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1913                                 pr_err("Failed to set freeze_on_smi.\n");
1914                                 ret = -1;
1915                                 goto out;
1916                         }
1917                         smi_reset = true;
1918                 }
1919
1920                 if (!metricgroup__has_metric(pmu, "smi")) {
1921                         pr_err("Missing smi metrics\n");
1922                         ret = -1;
1923                         goto out;
1924                 }
1925
1926                 if (!force_metric_only)
1927                         stat_config.metric_only = true;
1928
1929                 ret = metricgroup__parse_groups(evlist, pmu, "smi",
1930                                                 stat_config.metric_no_group,
1931                                                 stat_config.metric_no_merge,
1932                                                 stat_config.metric_no_threshold,
1933                                                 stat_config.user_requested_cpu_list,
1934                                                 stat_config.system_wide,
1935                                                 stat_config.hardware_aware_grouping,
1936                                                 &stat_config.metric_events);
1937                 goto out;
1938         }
1939
1940         if (topdown_run) {
1941                 unsigned int max_level = metricgroups__topdown_max_level();
1942                 char str[] = "TopdownL1";
1943
1944                 if (!force_metric_only)
1945                         stat_config.metric_only = true;
1946
1947                 if (!max_level) {
1948                         pr_err("Topdown requested but the topdown metric groups aren't present.\n"
1949                                 "(See perf list the metric groups have names like TopdownL1)\n");
1950                         ret = -1;
1951                         goto out;
1952                 }
1953                 if (stat_config.topdown_level > max_level) {
1954                         pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1955                         ret = -1;
1956                         goto out;
1957                 } else if (!stat_config.topdown_level) {
1958                         stat_config.topdown_level = 1;
1959                 }
1960                 if (!stat_config.interval && !stat_config.metric_only) {
1961                         fprintf(stat_config.output,
1962                                 "Topdown accuracy may decrease when measuring long periods.\n"
1963                                 "Please print the result regularly, e.g. -I1000\n");
1964                 }
1965                 str[8] = stat_config.topdown_level + '0';
1966                 if (metricgroup__parse_groups(evlist,
1967                                                 pmu, str,
1968                                                 /*metric_no_group=*/false,
1969                                                 /*metric_no_merge=*/false,
1970                                                 /*metric_no_threshold=*/true,
1971                                                 stat_config.user_requested_cpu_list,
1972                                                 stat_config.system_wide,
1973                                                 stat_config.hardware_aware_grouping,
1974                                                 &stat_config.metric_events) < 0) {
1975                         ret = -1;
1976                         goto out;
1977                 }
1978         }
1979
1980         if (!stat_config.topdown_level)
1981                 stat_config.topdown_level = 1;
1982
1983         if (!evlist->core.nr_entries && !evsel_list->core.nr_entries) {
1984                 /* No events so add defaults. */
1985                 if (target__has_cpu(&target))
1986                         ret = parse_events(evlist, "cpu-clock", &err);
1987                 else
1988                         ret = parse_events(evlist, "task-clock", &err);
1989                 if (ret)
1990                         goto out;
1991
1992                 ret = parse_events(evlist,
1993                                 "context-switches,"
1994                                 "cpu-migrations,"
1995                                 "page-faults,"
1996                                 "instructions,"
1997                                 "cycles,"
1998                                 "stalled-cycles-frontend,"
1999                                 "stalled-cycles-backend,"
2000                                 "branches,"
2001                                 "branch-misses",
2002                                 &err);
2003                 if (ret)
2004                         goto out;
2005
2006                 /*
2007                  * Add TopdownL1 metrics if they exist. To minimize
2008                  * multiplexing, don't request threshold computation.
2009                  */
2010                 if (metricgroup__has_metric(pmu, "Default")) {
2011                         struct evlist *metric_evlist = evlist__new();
2012
2013                         if (!metric_evlist) {
2014                                 ret = -ENOMEM;
2015                                 goto out;
2016                         }
2017                         if (metricgroup__parse_groups(metric_evlist, pmu, "Default",
2018                                                         /*metric_no_group=*/false,
2019                                                         /*metric_no_merge=*/false,
2020                                                         /*metric_no_threshold=*/true,
2021                                                         stat_config.user_requested_cpu_list,
2022                                                         stat_config.system_wide,
2023                                                         stat_config.hardware_aware_grouping,
2024                                                         &stat_config.metric_events) < 0) {
2025                                 ret = -1;
2026                                 goto out;
2027                         }
2028
2029                         evlist__for_each_entry(metric_evlist, evsel)
2030                                 evsel->default_metricgroup = true;
2031
2032                         evlist__splice_list_tail(evlist, &metric_evlist->core.entries);
2033                         evlist__delete(metric_evlist);
2034                 }
2035         }
2036
2037         /* Detailed events get appended to the event list: */
2038
2039         if (!ret && detailed_run >=  1) {
2040                 /*
2041                  * Detailed stats (-d), covering the L1 and last level data
2042                  * caches:
2043                  */
2044                 ret = parse_events(evlist,
2045                                 "L1-dcache-loads,"
2046                                 "L1-dcache-load-misses,"
2047                                 "LLC-loads,"
2048                                 "LLC-load-misses",
2049                                 &err);
2050         }
2051         if (!ret && detailed_run >=  2) {
2052                 /*
2053                  * Very detailed stats (-d -d), covering the instruction cache
2054                  * and the TLB caches:
2055                  */
2056                 ret = parse_events(evlist,
2057                                 "L1-icache-loads,"
2058                                 "L1-icache-load-misses,"
2059                                 "dTLB-loads,"
2060                                 "dTLB-load-misses,"
2061                                 "iTLB-loads,"
2062                                 "iTLB-load-misses",
2063                                 &err);
2064         }
2065         if (!ret && detailed_run >=  3) {
2066                 /*
2067                  * Very, very detailed stats (-d -d -d), adding prefetch events:
2068                  */
2069                 ret = parse_events(evlist,
2070                                 "L1-dcache-prefetches,"
2071                                 "L1-dcache-prefetch-misses",
2072                                 &err);
2073         }
2074 out:
2075         if (!ret) {
2076                 evlist__for_each_entry(evlist, evsel) {
2077                         /*
2078                          * Make at least one event non-skippable so fatal errors are visible.
2079                          * 'cycles' always used to be default and non-skippable, so use that.
2080                          */
2081                         if (strcmp("cycles", evsel__name(evsel)))
2082                                 evsel->skippable = true;
2083                 }
2084         }
2085         parse_events_error__exit(&err);
2086         evlist__splice_list_tail(evsel_list, &evlist->core.entries);
2087         evlist__delete(evlist);
2088         return ret;
2089 }
2090
2091 static const char * const stat_record_usage[] = {
2092         "perf stat record [<options>]",
2093         NULL,
2094 };
2095
2096 static void init_features(struct perf_session *session)
2097 {
2098         int feat;
2099
2100         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2101                 perf_header__set_feat(&session->header, feat);
2102
2103         perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2104         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2105         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2106         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2107         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2108 }
2109
2110 static int __cmd_record(const struct option stat_options[], struct opt_aggr_mode *opt_mode,
2111                         int argc, const char **argv)
2112 {
2113         struct perf_session *session;
2114         struct perf_data *data = &perf_stat.data;
2115
2116         argc = parse_options(argc, argv, stat_options, stat_record_usage,
2117                              PARSE_OPT_STOP_AT_NON_OPTION);
2118         stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(opt_mode);
2119
2120         if (output_name)
2121                 data->path = output_name;
2122
2123         if (stat_config.run_count != 1 || forever) {
2124                 pr_err("Cannot use -r option with perf stat record.\n");
2125                 return -1;
2126         }
2127
2128         session = perf_session__new(data, NULL);
2129         if (IS_ERR(session)) {
2130                 pr_err("Perf session creation failed\n");
2131                 return PTR_ERR(session);
2132         }
2133
2134         init_features(session);
2135
2136         session->evlist   = evsel_list;
2137         perf_stat.session = session;
2138         perf_stat.record  = true;
2139         return argc;
2140 }
2141
2142 static int process_stat_round_event(struct perf_session *session,
2143                                     union perf_event *event)
2144 {
2145         struct perf_record_stat_round *stat_round = &event->stat_round;
2146         struct timespec tsh, *ts = NULL;
2147         const char **argv = session->header.env.cmdline_argv;
2148         int argc = session->header.env.nr_cmdline;
2149
2150         process_counters();
2151
2152         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2153                 update_stats(&walltime_nsecs_stats, stat_round->time);
2154
2155         if (stat_config.interval && stat_round->time) {
2156                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2157                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2158                 ts = &tsh;
2159         }
2160
2161         print_counters(ts, argc, argv);
2162         return 0;
2163 }
2164
2165 static
2166 int process_stat_config_event(struct perf_session *session,
2167                               union perf_event *event)
2168 {
2169         const struct perf_tool *tool = session->tool;
2170         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2171
2172         perf_event__read_stat_config(&stat_config, &event->stat_config);
2173
2174         if (perf_cpu_map__is_empty(st->cpus)) {
2175                 if (st->aggr_mode != AGGR_UNSET)
2176                         pr_warning("warning: processing task data, aggregation mode not set\n");
2177         } else if (st->aggr_mode != AGGR_UNSET) {
2178                 stat_config.aggr_mode = st->aggr_mode;
2179         }
2180
2181         if (perf_stat.data.is_pipe)
2182                 perf_stat_init_aggr_mode();
2183         else
2184                 perf_stat_init_aggr_mode_file(st);
2185
2186         if (stat_config.aggr_map) {
2187                 int nr_aggr = stat_config.aggr_map->nr;
2188
2189                 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2190                         pr_err("cannot allocate aggr counts\n");
2191                         return -1;
2192                 }
2193         }
2194         return 0;
2195 }
2196
2197 static int set_maps(struct perf_stat *st)
2198 {
2199         if (!st->cpus || !st->threads)
2200                 return 0;
2201
2202         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2203                 return -EINVAL;
2204
2205         perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2206
2207         if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2208                 return -ENOMEM;
2209
2210         st->maps_allocated = true;
2211         return 0;
2212 }
2213
2214 static
2215 int process_thread_map_event(struct perf_session *session,
2216                              union perf_event *event)
2217 {
2218         const struct perf_tool *tool = session->tool;
2219         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2220
2221         if (st->threads) {
2222                 pr_warning("Extra thread map event, ignoring.\n");
2223                 return 0;
2224         }
2225
2226         st->threads = thread_map__new_event(&event->thread_map);
2227         if (!st->threads)
2228                 return -ENOMEM;
2229
2230         return set_maps(st);
2231 }
2232
2233 static
2234 int process_cpu_map_event(struct perf_session *session,
2235                           union perf_event *event)
2236 {
2237         const struct perf_tool *tool = session->tool;
2238         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2239         struct perf_cpu_map *cpus;
2240
2241         if (st->cpus) {
2242                 pr_warning("Extra cpu map event, ignoring.\n");
2243                 return 0;
2244         }
2245
2246         cpus = cpu_map__new_data(&event->cpu_map.data);
2247         if (!cpus)
2248                 return -ENOMEM;
2249
2250         st->cpus = cpus;
2251         return set_maps(st);
2252 }
2253
2254 static const char * const stat_report_usage[] = {
2255         "perf stat report [<options>]",
2256         NULL,
2257 };
2258
2259 static struct perf_stat perf_stat = {
2260         .aggr_mode      = AGGR_UNSET,
2261         .aggr_level     = 0,
2262 };
2263
2264 static int __cmd_report(int argc, const char **argv)
2265 {
2266         struct perf_session *session;
2267         const struct option options[] = {
2268         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2269         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2270                      "aggregate counts per processor socket", AGGR_SOCKET),
2271         OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2272                      "aggregate counts per processor die", AGGR_DIE),
2273         OPT_SET_UINT(0, "per-cluster", &perf_stat.aggr_mode,
2274                      "aggregate counts perf processor cluster", AGGR_CLUSTER),
2275         OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
2276                             "cache level",
2277                             "aggregate count at this cache level (Default: LLC)",
2278                             parse_cache_level),
2279         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2280                      "aggregate counts per physical processor core", AGGR_CORE),
2281         OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2282                      "aggregate counts per numa node", AGGR_NODE),
2283         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2284                      "disable CPU count aggregation", AGGR_NONE),
2285         OPT_END()
2286         };
2287         struct stat st;
2288         int ret;
2289
2290         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2291
2292         if (!input_name || !strlen(input_name)) {
2293                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2294                         input_name = "-";
2295                 else
2296                         input_name = "perf.data";
2297         }
2298
2299         perf_stat.data.path = input_name;
2300         perf_stat.data.mode = PERF_DATA_MODE_READ;
2301
2302         perf_tool__init(&perf_stat.tool, /*ordered_events=*/false);
2303         perf_stat.tool.attr             = perf_event__process_attr;
2304         perf_stat.tool.event_update     = perf_event__process_event_update;
2305         perf_stat.tool.thread_map       = process_thread_map_event;
2306         perf_stat.tool.cpu_map          = process_cpu_map_event;
2307         perf_stat.tool.stat_config      = process_stat_config_event;
2308         perf_stat.tool.stat             = perf_event__process_stat_event;
2309         perf_stat.tool.stat_round       = process_stat_round_event;
2310
2311         session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2312         if (IS_ERR(session))
2313                 return PTR_ERR(session);
2314
2315         perf_stat.session  = session;
2316         stat_config.output = stderr;
2317         evlist__delete(evsel_list);
2318         evsel_list         = session->evlist;
2319
2320         ret = perf_session__process_events(session);
2321         if (ret)
2322                 return ret;
2323
2324         perf_session__delete(session);
2325         return 0;
2326 }
2327
2328 static void setup_system_wide(int forks)
2329 {
2330         /*
2331          * Make system wide (-a) the default target if
2332          * no target was specified and one of following
2333          * conditions is met:
2334          *
2335          *   - there's no workload specified
2336          *   - there is workload specified but all requested
2337          *     events are system wide events
2338          */
2339         if (!target__none(&target))
2340                 return;
2341
2342         if (!forks)
2343                 target.system_wide = true;
2344         else {
2345                 struct evsel *counter;
2346
2347                 evlist__for_each_entry(evsel_list, counter) {
2348                         if (!counter->core.requires_cpu &&
2349                             !evsel__name_is(counter, "duration_time")) {
2350                                 return;
2351                         }
2352                 }
2353
2354                 if (evsel_list->core.nr_entries)
2355                         target.system_wide = true;
2356         }
2357 }
2358
2359 int cmd_stat(int argc, const char **argv)
2360 {
2361         struct opt_aggr_mode opt_mode = {};
2362         struct option stat_options[] = {
2363                 OPT_BOOLEAN('T', "transaction", &transaction_run,
2364                         "hardware transaction statistics"),
2365                 OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
2366                         "event selector. use 'perf list' to list available events",
2367                         parse_events_option),
2368                 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
2369                         "event filter", parse_filter),
2370                 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
2371                         "child tasks do not inherit counters"),
2372                 OPT_STRING('p', "pid", &target.pid, "pid",
2373                         "stat events on existing process id"),
2374                 OPT_STRING('t', "tid", &target.tid, "tid",
2375                         "stat events on existing thread id"),
2376 #ifdef HAVE_BPF_SKEL
2377                 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
2378                         "stat events on existing bpf program id"),
2379                 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
2380                         "use bpf program to count events"),
2381                 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
2382                         "path to perf_event_attr map"),
2383 #endif
2384                 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
2385                         "system-wide collection from all CPUs"),
2386                 OPT_BOOLEAN(0, "scale", &stat_config.scale,
2387                         "Use --no-scale to disable counter scaling for multiplexing"),
2388                 OPT_INCR('v', "verbose", &verbose,
2389                         "be more verbose (show counter open errors, etc)"),
2390                 OPT_INTEGER('r', "repeat", &stat_config.run_count,
2391                         "repeat command and print average + stddev (max: 100, forever: 0)"),
2392                 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
2393                         "display details about each run (only with -r option)"),
2394                 OPT_BOOLEAN('n', "null", &stat_config.null_run,
2395                         "null run - dont start any counters"),
2396                 OPT_INCR('d', "detailed", &detailed_run,
2397                         "detailed run - start a lot of events"),
2398                 OPT_BOOLEAN('S', "sync", &sync_run,
2399                         "call sync() before starting a run"),
2400                 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
2401                                 "print large numbers with thousands\' separators",
2402                                 stat__set_big_num),
2403                 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
2404                         "list of cpus to monitor in system-wide"),
2405                 OPT_BOOLEAN('A', "no-aggr", &opt_mode.no_aggr,
2406                         "disable aggregation across CPUs or PMUs"),
2407                 OPT_BOOLEAN(0, "no-merge", &opt_mode.no_aggr,
2408                         "disable aggregation the same as -A or -no-aggr"),
2409                 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
2410                         "Merge identical named hybrid events"),
2411                 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
2412                         "print counts with custom separator"),
2413                 OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
2414                         "print counts in JSON format"),
2415                 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
2416                         "monitor event in cgroup name only", parse_stat_cgroups),
2417                 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
2418                         "expand events for each cgroup"),
2419                 OPT_STRING('o', "output", &output_name, "file", "output file name"),
2420                 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
2421                 OPT_INTEGER(0, "log-fd", &output_fd,
2422                         "log output to fd, instead of stderr"),
2423                 OPT_STRING(0, "pre", &pre_cmd, "command",
2424                         "command to run prior to the measured command"),
2425                 OPT_STRING(0, "post", &post_cmd, "command",
2426                         "command to run after to the measured command"),
2427                 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
2428                         "print counts at regular interval in ms "
2429                         "(overhead is possible for values <= 100ms)"),
2430                 OPT_INTEGER(0, "interval-count", &stat_config.times,
2431                         "print counts for fixed number of times"),
2432                 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
2433                         "clear screen in between new interval"),
2434                 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
2435                         "stop workload and print counts after a timeout period in ms (>= 10ms)"),
2436                 OPT_BOOLEAN(0, "per-socket", &opt_mode.socket,
2437                         "aggregate counts per processor socket"),
2438                 OPT_BOOLEAN(0, "per-die", &opt_mode.die, "aggregate counts per processor die"),
2439                 OPT_BOOLEAN(0, "per-cluster", &opt_mode.cluster,
2440                         "aggregate counts per processor cluster"),
2441                 OPT_CALLBACK_OPTARG(0, "per-cache", &opt_mode, &stat_config.aggr_level,
2442                                 "cache level", "aggregate count at this cache level (Default: LLC)",
2443                                 parse_cache_level),
2444                 OPT_BOOLEAN(0, "per-core", &opt_mode.core,
2445                         "aggregate counts per physical processor core"),
2446                 OPT_BOOLEAN(0, "per-thread", &opt_mode.thread, "aggregate counts per thread"),
2447                 OPT_BOOLEAN(0, "per-node", &opt_mode.node, "aggregate counts per numa node"),
2448                 OPT_INTEGER('D', "delay", &target.initial_delay,
2449                         "ms to wait before starting measurement after program start (-1: start with events disabled)"),
2450                 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
2451                                 "Only print computed metrics. No raw values", enable_metric_only),
2452                 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
2453                         "don't group metric events, impacts multiplexing"),
2454                 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
2455                         "don't try to share events between metrics in a group"),
2456                 OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold,
2457                         "disable adding events for the metric threshold calculation"),
2458                 OPT_BOOLEAN(0, "topdown", &topdown_run,
2459                         "measure top-down statistics"),
2460 #ifdef HAVE_ARCH_X86_64_SUPPORT
2461                 OPT_BOOLEAN(0, "record-tpebs", &tpebs_recording,
2462                         "enable recording for tpebs when retire_latency required"),
2463 #endif
2464                 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
2465                         "Set the metrics level for the top-down statistics (0: max level)"),
2466                 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
2467                         "measure SMI cost"),
2468                 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
2469                         "monitor specified metrics or metric groups (separated by ,)",
2470                         append_metric_groups),
2471                 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
2472                                 "Configure all used events to run in kernel space.",
2473                                 PARSE_OPT_EXCLUSIVE),
2474                 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
2475                                 "Configure all used events to run in user space.",
2476                                 PARSE_OPT_EXCLUSIVE),
2477                 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
2478                         "Use with 'percore' event qualifier to show the event "
2479                         "counts of one hardware thread by sum up total hardware "
2480                         "threads of same physical core"),
2481                 OPT_BOOLEAN(0, "summary", &stat_config.summary,
2482                         "print summary for interval mode"),
2483                 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
2484                         "don't print 'summary' for CSV summary output"),
2485                 OPT_BOOLEAN(0, "quiet", &quiet,
2486                         "don't print any output, messages or warnings (useful with record)"),
2487                 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
2488                         "Only enable events on applying cpu with this type "
2489                         "for hybrid platform (e.g. core or atom)",
2490                         parse_cputype),
2491 #ifdef HAVE_LIBPFM
2492                 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
2493                         "libpfm4 event selector. use 'perf list' to list available events",
2494                         parse_libpfm_events_option),
2495 #endif
2496                 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
2497                         "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
2498                         "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
2499                         "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
2500                         parse_control_option),
2501                 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
2502                                 "measure I/O performance metrics provided by arch/platform",
2503                                 iostat_parse),
2504                 OPT_END()
2505         };
2506         const char * const stat_usage[] = {
2507                 "perf stat [<options>] [<command>]",
2508                 NULL
2509         };
2510         int status = -EINVAL, run_idx, err;
2511         const char *mode;
2512         FILE *output = stderr;
2513         unsigned int interval, timeout;
2514         const char * const stat_subcommands[] = { "record", "report" };
2515         char errbuf[BUFSIZ];
2516
2517         setlocale(LC_ALL, "");
2518
2519         evsel_list = evlist__new();
2520         if (evsel_list == NULL)
2521                 return -ENOMEM;
2522
2523         parse_events__shrink_config_terms();
2524
2525         /* String-parsing callback-based options would segfault when negated */
2526         set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2527         set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2528         set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2529
2530         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2531                                         (const char **) stat_usage,
2532                                         PARSE_OPT_STOP_AT_NON_OPTION);
2533
2534         stat_config.aggr_mode = opt_aggr_mode_to_aggr_mode(&opt_mode);
2535
2536         if (stat_config.csv_sep) {
2537                 stat_config.csv_output = true;
2538                 if (!strcmp(stat_config.csv_sep, "\\t"))
2539                         stat_config.csv_sep = "\t";
2540         } else
2541                 stat_config.csv_sep = DEFAULT_SEPARATOR;
2542
2543         if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2544                 argc = __cmd_record(stat_options, &opt_mode, argc, argv);
2545                 if (argc < 0)
2546                         return -1;
2547         } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2548                 return __cmd_report(argc, argv);
2549
2550         interval = stat_config.interval;
2551         timeout = stat_config.timeout;
2552
2553         /*
2554          * For record command the -o is already taken care of.
2555          */
2556         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2557                 output = NULL;
2558
2559         if (output_name && output_fd) {
2560                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2561                 parse_options_usage(stat_usage, stat_options, "o", 1);
2562                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2563                 goto out;
2564         }
2565
2566         if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2567                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2568                 goto out;
2569         }
2570
2571         if (stat_config.metric_only && stat_config.run_count > 1) {
2572                 fprintf(stderr, "--metric-only is not supported with -r\n");
2573                 goto out;
2574         }
2575
2576         if (stat_config.csv_output || (stat_config.metric_only && stat_config.json_output)) {
2577                 /*
2578                  * Current CSV and metric-only JSON output doesn't display the
2579                  * metric threshold so don't compute it.
2580                  */
2581                 stat_config.metric_no_threshold = true;
2582         }
2583
2584         if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2585                 fprintf(stderr, "--table is only supported with -r\n");
2586                 parse_options_usage(stat_usage, stat_options, "r", 1);
2587                 parse_options_usage(NULL, stat_options, "table", 0);
2588                 goto out;
2589         }
2590
2591         if (output_fd < 0) {
2592                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2593                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2594                 goto out;
2595         }
2596
2597         if (!output && !quiet) {
2598                 struct timespec tm;
2599                 mode = append_file ? "a" : "w";
2600
2601                 output = fopen(output_name, mode);
2602                 if (!output) {
2603                         perror("failed to create output file");
2604                         return -1;
2605                 }
2606                 if (!stat_config.json_output) {
2607                         clock_gettime(CLOCK_REALTIME, &tm);
2608                         fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2609                 }
2610         } else if (output_fd > 0) {
2611                 mode = append_file ? "a" : "w";
2612                 output = fdopen(output_fd, mode);
2613                 if (!output) {
2614                         perror("Failed opening logfd");
2615                         return -errno;
2616                 }
2617         }
2618
2619         if (stat_config.interval_clear && !isatty(fileno(output))) {
2620                 fprintf(stderr, "--interval-clear does not work with output\n");
2621                 parse_options_usage(stat_usage, stat_options, "o", 1);
2622                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2623                 parse_options_usage(NULL, stat_options, "interval-clear", 0);
2624                 return -1;
2625         }
2626
2627         stat_config.output = output;
2628
2629         /*
2630          * let the spreadsheet do the pretty-printing
2631          */
2632         if (stat_config.csv_output) {
2633                 /* User explicitly passed -B? */
2634                 if (big_num_opt == 1) {
2635                         fprintf(stderr, "-B option not supported with -x\n");
2636                         parse_options_usage(stat_usage, stat_options, "B", 1);
2637                         parse_options_usage(NULL, stat_options, "x", 1);
2638                         goto out;
2639                 } else /* Nope, so disable big number formatting */
2640                         stat_config.big_num = false;
2641         } else if (big_num_opt == 0) /* User passed --no-big-num */
2642                 stat_config.big_num = false;
2643
2644         target.inherit = !stat_config.no_inherit;
2645         err = target__validate(&target);
2646         if (err) {
2647                 target__strerror(&target, err, errbuf, BUFSIZ);
2648                 pr_warning("%s\n", errbuf);
2649         }
2650
2651         setup_system_wide(argc);
2652
2653         /*
2654          * Display user/system times only for single
2655          * run and when there's specified tracee.
2656          */
2657         if ((stat_config.run_count == 1) && target__none(&target))
2658                 stat_config.ru_display = true;
2659
2660         if (stat_config.run_count < 0) {
2661                 pr_err("Run count must be a positive number\n");
2662                 parse_options_usage(stat_usage, stat_options, "r", 1);
2663                 goto out;
2664         } else if (stat_config.run_count == 0) {
2665                 forever = true;
2666                 stat_config.run_count = 1;
2667         }
2668
2669         if (stat_config.walltime_run_table) {
2670                 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2671                 if (!stat_config.walltime_run) {
2672                         pr_err("failed to setup -r option");
2673                         goto out;
2674                 }
2675         }
2676
2677         if ((stat_config.aggr_mode == AGGR_THREAD) &&
2678                 !target__has_task(&target)) {
2679                 if (!target.system_wide || target.cpu_list) {
2680                         fprintf(stderr, "The --per-thread option is only "
2681                                 "available when monitoring via -p -t -a "
2682                                 "options or only --per-thread.\n");
2683                         parse_options_usage(NULL, stat_options, "p", 1);
2684                         parse_options_usage(NULL, stat_options, "t", 1);
2685                         goto out;
2686                 }
2687         }
2688
2689         /*
2690          * no_aggr, cgroup are for system-wide only
2691          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2692          */
2693         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2694               stat_config.aggr_mode != AGGR_THREAD) ||
2695              (nr_cgroups || stat_config.cgroup_list)) &&
2696             !target__has_cpu(&target)) {
2697                 fprintf(stderr, "both cgroup and no-aggregation "
2698                         "modes only available in system-wide mode\n");
2699
2700                 parse_options_usage(stat_usage, stat_options, "G", 1);
2701                 parse_options_usage(NULL, stat_options, "A", 1);
2702                 parse_options_usage(NULL, stat_options, "a", 1);
2703                 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2704                 goto out;
2705         }
2706
2707         if (stat_config.iostat_run) {
2708                 status = iostat_prepare(evsel_list, &stat_config);
2709                 if (status)
2710                         goto out;
2711                 if (iostat_mode == IOSTAT_LIST) {
2712                         iostat_list(evsel_list, &stat_config);
2713                         goto out;
2714                 } else if (verbose > 0)
2715                         iostat_list(evsel_list, &stat_config);
2716                 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2717                         target.system_wide = true;
2718         }
2719
2720         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2721                 target.per_thread = true;
2722
2723         stat_config.system_wide = target.system_wide;
2724         if (target.cpu_list) {
2725                 stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2726                 if (!stat_config.user_requested_cpu_list) {
2727                         status = -ENOMEM;
2728                         goto out;
2729                 }
2730         }
2731
2732         /*
2733          * Metric parsing needs to be delayed as metrics may optimize events
2734          * knowing the target is system-wide.
2735          */
2736         if (metrics) {
2737                 const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2738                 int ret = metricgroup__parse_groups(evsel_list, pmu, metrics,
2739                                                 stat_config.metric_no_group,
2740                                                 stat_config.metric_no_merge,
2741                                                 stat_config.metric_no_threshold,
2742                                                 stat_config.user_requested_cpu_list,
2743                                                 stat_config.system_wide,
2744                                                 stat_config.hardware_aware_grouping,
2745                                                 &stat_config.metric_events);
2746
2747                 zfree(&metrics);
2748                 if (ret) {
2749                         status = ret;
2750                         goto out;
2751                 }
2752         }
2753
2754         if (add_default_events())
2755                 goto out;
2756
2757         if (stat_config.cgroup_list) {
2758                 if (nr_cgroups > 0) {
2759                         pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2760                         parse_options_usage(stat_usage, stat_options, "G", 1);
2761                         parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2762                         goto out;
2763                 }
2764
2765                 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2766                                           &stat_config.metric_events, true) < 0) {
2767                         parse_options_usage(stat_usage, stat_options,
2768                                             "for-each-cgroup", 0);
2769                         goto out;
2770                 }
2771         }
2772
2773         evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
2774
2775         if (evlist__create_maps(evsel_list, &target) < 0) {
2776                 if (target__has_task(&target)) {
2777                         pr_err("Problems finding threads of monitor\n");
2778                         parse_options_usage(stat_usage, stat_options, "p", 1);
2779                         parse_options_usage(NULL, stat_options, "t", 1);
2780                 } else if (target__has_cpu(&target)) {
2781                         perror("failed to parse CPUs map");
2782                         parse_options_usage(stat_usage, stat_options, "C", 1);
2783                         parse_options_usage(NULL, stat_options, "a", 1);
2784                 }
2785                 goto out;
2786         }
2787
2788         evlist__check_cpu_maps(evsel_list);
2789
2790         /*
2791          * Initialize thread_map with comm names,
2792          * so we could print it out on output.
2793          */
2794         if (stat_config.aggr_mode == AGGR_THREAD) {
2795                 thread_map__read_comms(evsel_list->core.threads);
2796         }
2797
2798         if (stat_config.aggr_mode == AGGR_NODE)
2799                 cpu__setup_cpunode_map();
2800
2801         if (stat_config.times && interval)
2802                 interval_count = true;
2803         else if (stat_config.times && !interval) {
2804                 pr_err("interval-count option should be used together with "
2805                                 "interval-print.\n");
2806                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2807                 parse_options_usage(stat_usage, stat_options, "I", 1);
2808                 goto out;
2809         }
2810
2811         if (timeout && timeout < 100) {
2812                 if (timeout < 10) {
2813                         pr_err("timeout must be >= 10ms.\n");
2814                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
2815                         goto out;
2816                 } else
2817                         pr_warning("timeout < 100ms. "
2818                                    "The overhead percentage could be high in some cases. "
2819                                    "Please proceed with caution.\n");
2820         }
2821         if (timeout && interval) {
2822                 pr_err("timeout option is not supported with interval-print.\n");
2823                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2824                 parse_options_usage(stat_usage, stat_options, "I", 1);
2825                 goto out;
2826         }
2827
2828         if (perf_stat_init_aggr_mode())
2829                 goto out;
2830
2831         if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2832                 goto out;
2833
2834         /*
2835          * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2836          * while avoiding that older tools show confusing messages.
2837          *
2838          * However for pipe sessions we need to keep it zero,
2839          * because script's perf_evsel__check_attr is triggered
2840          * by attr->sample_type != 0, and we can't run it on
2841          * stat sessions.
2842          */
2843         stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2844
2845         /*
2846          * We dont want to block the signals - that would cause
2847          * child tasks to inherit that and Ctrl-C would not work.
2848          * What we want is for Ctrl-C to work in the exec()-ed
2849          * task, but being ignored by perf stat itself:
2850          */
2851         atexit(sig_atexit);
2852         if (!forever)
2853                 signal(SIGINT,  skip_signal);
2854         signal(SIGCHLD, skip_signal);
2855         signal(SIGALRM, skip_signal);
2856         signal(SIGABRT, skip_signal);
2857
2858         if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2859                 goto out;
2860
2861         /* Enable ignoring missing threads when -p option is defined. */
2862         evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2863         status = 0;
2864         for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2865                 if (stat_config.run_count != 1 && verbose > 0)
2866                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2867                                 run_idx + 1);
2868
2869                 if (run_idx != 0)
2870                         evlist__reset_prev_raw_counts(evsel_list);
2871
2872                 status = run_perf_stat(argc, argv, run_idx);
2873                 if (status == -1)
2874                         break;
2875
2876                 if (forever && !interval) {
2877                         print_counters(NULL, argc, argv);
2878                         perf_stat__reset_stats();
2879                 }
2880         }
2881
2882         if (!forever && status != -1 && (!interval || stat_config.summary)) {
2883                 if (stat_config.run_count > 1)
2884                         evlist__copy_res_stats(&stat_config, evsel_list);
2885                 print_counters(NULL, argc, argv);
2886         }
2887
2888         evlist__finalize_ctlfd(evsel_list);
2889
2890         if (STAT_RECORD) {
2891                 /*
2892                  * We synthesize the kernel mmap record just so that older tools
2893                  * don't emit warnings about not being able to resolve symbols
2894                  * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2895                  * a saner message about no samples being in the perf.data file.
2896                  *
2897                  * This also serves to suppress a warning about f_header.data.size == 0
2898                  * in header.c at the moment 'perf stat record' gets introduced, which
2899                  * is not really needed once we start adding the stat specific PERF_RECORD_
2900                  * records, but the need to suppress the kptr_restrict messages in older
2901                  * tools remain  -acme
2902                  */
2903                 int fd = perf_data__fd(&perf_stat.data);
2904
2905                 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2906                                                          process_synthesized_event,
2907                                                          &perf_stat.session->machines.host);
2908                 if (err) {
2909                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2910                                    "older tools may produce warnings about this file\n.");
2911                 }
2912
2913                 if (!interval) {
2914                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2915                                 pr_err("failed to write stat round event\n");
2916                 }
2917
2918                 if (!perf_stat.data.is_pipe) {
2919                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2920                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2921                 }
2922
2923                 evlist__close(evsel_list);
2924                 perf_session__delete(perf_stat.session);
2925         }
2926
2927         perf_stat__exit_aggr_mode();
2928         evlist__free_stats(evsel_list);
2929 out:
2930         if (stat_config.iostat_run)
2931                 iostat_release(evsel_list);
2932
2933         zfree(&stat_config.walltime_run);
2934         zfree(&stat_config.user_requested_cpu_list);
2935
2936         if (smi_cost && smi_reset)
2937                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2938
2939         evlist__delete(evsel_list);
2940
2941         metricgroup__rblist_exit(&stat_config.metric_events);
2942         evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2943
2944         return status;
2945 }