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/event.h"
  50 #include "util/evlist.h"
  51 #include "util/evsel.h"
  52 #include "util/debug.h"
  53 #include "util/color.h"
  54 #include "util/stat.h"
  55 #include "util/header.h"
  56 #include "util/cpumap.h"
  57 #include "util/thread_map.h"
  58 #include "util/counts.h"
  59 #include "util/topdown.h"
  60 #include "util/session.h"
  61 #include "util/tool.h"
  62 #include "util/string2.h"
  63 #include "util/metricgroup.h"
  64 #include "util/synthetic-events.h"
  65 #include "util/target.h"
  66 #include "util/time-utils.h"
  67 #include "util/top.h"
  68 #include "util/affinity.h"
  69 #include "util/pfm.h"
  70 #include "util/bpf_counter.h"
  71 #include "util/iostat.h"
  72 #include "util/util.h"
  73 #include "asm/bug.h"
  74
  75 #include <linux/time64.h>
  76 #include <linux/zalloc.h>
  77 #include <api/fs/fs.h>
  78 #include <errno.h>
  79 #include <signal.h>
  80 #include <stdlib.h>
  81 #include <sys/prctl.h>
  82 #include <inttypes.h>
  83 #include <locale.h>
  84 #include <math.h>
  85 #include <sys/types.h>
  86 #include <sys/stat.h>
  87 #include <sys/wait.h>
  88 #include <unistd.h>
  89 #include <sys/time.h>
  90 #include <sys/resource.h>
  91 #include <linux/err.h>
  92
  93 #include <linux/ctype.h>
  94 #include <perf/evlist.h>
  95 #include <internal/threadmap.h>
  96
  97 #define DEFAULT_SEPARATOR       " "
  98 #define FREEZE_ON_SMI_PATH      "devices/cpu/freeze_on_smi"
  99
 100 static void print_counters(struct timespec *ts, int argc, const char **argv);
 101
 102 static struct evlist    *evsel_list;
 103 static struct parse_events_option_args parse_events_option_args = {
 104         .evlistp = &evsel_list,
 105 };
 106
 107 static bool all_counters_use_bpf = true;
 108
 109 static struct target target = {
 110         .uid    = UINT_MAX,
 111 };
 112
 113 #define METRIC_ONLY_LEN 20
 114
 115 static volatile sig_atomic_t    child_pid                       = -1;
 116 static int                      detailed_run                    =  0;
 117 static bool                     transaction_run;
 118 static bool                     topdown_run                     = false;
 119 static bool                     smi_cost                        = false;
 120 static bool                     smi_reset                       = false;
 121 static int                      big_num_opt                     =  -1;
 122 static const char               *pre_cmd                        = NULL;
 123 static const char               *post_cmd                       = NULL;
 124 static bool                     sync_run                        = false;
 125 static bool                     forever                         = false;
 126 static bool                     force_metric_only               = false;
 127 static struct timespec          ref_time;
 128 static bool                     append_file;
 129 static bool                     interval_count;
 130 static const char               *output_name;
 131 static int                      output_fd;
 132 static char                     *metrics;
 133
 134 struct perf_stat {
 135         bool                     record;
 136         struct perf_data         data;
 137         struct perf_session     *session;
 138         u64                      bytes_written;
 139         struct perf_tool         tool;
 140         bool                     maps_allocated;
 141         struct perf_cpu_map     *cpus;
 142         struct perf_thread_map *threads;
 143         enum aggr_mode           aggr_mode;
 144         u32                      aggr_level;
 145 };
 146
 147 static struct perf_stat         perf_stat;
 148 #define STAT_RECORD             perf_stat.record
 149
 150 static volatile sig_atomic_t done = 0;
 151
 152 static struct perf_stat_config stat_config = {
 153         .aggr_mode              = AGGR_GLOBAL,
 154         .aggr_level             = MAX_CACHE_LVL + 1,
 155         .scale                  = true,
 156         .unit_width             = 4, /* strlen("unit") */
 157         .run_count              = 1,
 158         .metric_only_len        = METRIC_ONLY_LEN,
 159         .walltime_nsecs_stats   = &walltime_nsecs_stats,
 160         .ru_stats               = &ru_stats,
 161         .big_num                = true,
 162         .ctl_fd                 = -1,
 163         .ctl_fd_ack             = -1,
 164         .iostat_run             = false,
 165 };
 166
 167 static void evlist__check_cpu_maps(struct evlist *evlist)
 168 {
 169         struct evsel *evsel, *warned_leader = NULL;
 170
 171         evlist__for_each_entry(evlist, evsel) {
 172                 struct evsel *leader = evsel__leader(evsel);
 173
 174                 /* Check that leader matches cpus with each member. */
 175                 if (leader == evsel)
 176                         continue;
 177                 if (perf_cpu_map__equal(leader->core.cpus, evsel->core.cpus))
 178                         continue;
 179
 180                 /* If there's mismatch disable the group and warn user. */
 181                 if (warned_leader != leader) {
 182                         char buf[200];
 183
 184                         pr_warning("WARNING: grouped events cpus do not match.\n"
 185                                 "Events with CPUs not matching the leader will "
 186                                 "be removed from the group.\n");
 187                         evsel__group_desc(leader, buf, sizeof(buf));
 188                         pr_warning("  %s\n", buf);
 189                         warned_leader = leader;
 190                 }
 191                 if (verbose > 0) {
 192                         char buf[200];
 193
 194                         cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
 195                         pr_warning("     %s: %s\n", leader->name, buf);
 196                         cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
 197                         pr_warning("     %s: %s\n", evsel->name, buf);
 198                 }
 199
 200                 evsel__remove_from_group(evsel, leader);
 201         }
 202 }
 203
 204 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 205                                  struct timespec *b)
 206 {
 207         r->tv_sec = a->tv_sec - b->tv_sec;
 208         if (a->tv_nsec < b->tv_nsec) {
 209                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 210                 r->tv_sec--;
 211         } else {
 212                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 213         }
 214 }
 215
 216 static void perf_stat__reset_stats(void)
 217 {
 218         evlist__reset_stats(evsel_list);
 219         perf_stat__reset_shadow_stats();
 220 }
 221
 222 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 223                                      union perf_event *event,
 224                                      struct perf_sample *sample __maybe_unused,
 225                                      struct machine *machine __maybe_unused)
 226 {
 227         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 228                 pr_err("failed to write perf data, error: %m\n");
 229                 return -1;
 230         }
 231
 232         perf_stat.bytes_written += event->header.size;
 233         return 0;
 234 }
 235
 236 static int write_stat_round_event(u64 tm, u64 type)
 237 {
 238         return perf_event__synthesize_stat_round(NULL, tm, type,
 239                                                  process_synthesized_event,
 240                                                  NULL);
 241 }
 242
 243 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 244         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 245
 246 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
 247
 248 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
 249                                    struct perf_counts_values *count)
 250 {
 251         struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
 252         struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
 253
 254         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 255                                            process_synthesized_event, NULL);
 256 }
 257
 258 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
 259                                int thread, struct timespec *rs)
 260 {
 261         switch(counter->tool_event) {
 262                 case PERF_TOOL_DURATION_TIME: {
 263                         u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
 264                         struct perf_counts_values *count =
 265                                 perf_counts(counter->counts, cpu_map_idx, thread);
 266                         count->ena = count->run = val;
 267                         count->val = val;
 268                         return 0;
 269                 }
 270                 case PERF_TOOL_USER_TIME:
 271                 case PERF_TOOL_SYSTEM_TIME: {
 272                         u64 val;
 273                         struct perf_counts_values *count =
 274                                 perf_counts(counter->counts, cpu_map_idx, thread);
 275                         if (counter->tool_event == PERF_TOOL_USER_TIME)
 276                                 val = ru_stats.ru_utime_usec_stat.mean;
 277                         else
 278                                 val = ru_stats.ru_stime_usec_stat.mean;
 279                         count->ena = count->run = val;
 280                         count->val = val;
 281                         return 0;
 282                 }
 283                 default:
 284                 case PERF_TOOL_NONE:
 285                         return evsel__read_counter(counter, cpu_map_idx, thread);
 286                 case PERF_TOOL_MAX:
 287                         /* This should never be reached */
 288                         return 0;
 289         }
 290 }
 291
 292 /*
 293  * Read out the results of a single counter:
 294  * do not aggregate counts across CPUs in system-wide mode
 295  */
 296 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
 297 {
 298         int nthreads = perf_thread_map__nr(evsel_list->core.threads);
 299         int thread;
 300
 301         if (!counter->supported)
 302                 return -ENOENT;
 303
 304         for (thread = 0; thread < nthreads; thread++) {
 305                 struct perf_counts_values *count;
 306
 307                 count = perf_counts(counter->counts, cpu_map_idx, thread);
 308
 309                 /*
 310                  * The leader's group read loads data into its group members
 311                  * (via evsel__read_counter()) and sets their count->loaded.
 312                  */
 313                 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
 314                     read_single_counter(counter, cpu_map_idx, thread, rs)) {
 315                         counter->counts->scaled = -1;
 316                         perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
 317                         perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
 318                         return -1;
 319                 }
 320
 321                 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
 322
 323                 if (STAT_RECORD) {
 324                         if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
 325                                 pr_err("failed to write stat event\n");
 326                                 return -1;
 327                         }
 328                 }
 329
 330                 if (verbose > 1) {
 331                         fprintf(stat_config.output,
 332                                 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 333                                         evsel__name(counter),
 334                                         perf_cpu_map__cpu(evsel__cpus(counter),
 335                                                           cpu_map_idx).cpu,
 336                                         count->val, count->ena, count->run);
 337                 }
 338         }
 339
 340         return 0;
 341 }
 342
 343 static int read_affinity_counters(struct timespec *rs)
 344 {
 345         struct evlist_cpu_iterator evlist_cpu_itr;
 346         struct affinity saved_affinity, *affinity;
 347
 348         if (all_counters_use_bpf)
 349                 return 0;
 350
 351         if (!target__has_cpu(&target) || target__has_per_thread(&target))
 352                 affinity = NULL;
 353         else if (affinity__setup(&saved_affinity) < 0)
 354                 return -1;
 355         else
 356                 affinity = &saved_affinity;
 357
 358         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 359                 struct evsel *counter = evlist_cpu_itr.evsel;
 360
 361                 if (evsel__is_bpf(counter))
 362                         continue;
 363
 364                 if (!counter->err) {
 365                         counter->err = read_counter_cpu(counter, rs,
 366                                                         evlist_cpu_itr.cpu_map_idx);
 367                 }
 368         }
 369         if (affinity)
 370                 affinity__cleanup(&saved_affinity);
 371
 372         return 0;
 373 }
 374
 375 static int read_bpf_map_counters(void)
 376 {
 377         struct evsel *counter;
 378         int err;
 379
 380         evlist__for_each_entry(evsel_list, counter) {
 381                 if (!evsel__is_bpf(counter))
 382                         continue;
 383
 384                 err = bpf_counter__read(counter);
 385                 if (err)
 386                         return err;
 387         }
 388         return 0;
 389 }
 390
 391 static int read_counters(struct timespec *rs)
 392 {
 393         if (!stat_config.stop_read_counter) {
 394                 if (read_bpf_map_counters() ||
 395                     read_affinity_counters(rs))
 396                         return -1;
 397         }
 398         return 0;
 399 }
 400
 401 static void process_counters(void)
 402 {
 403         struct evsel *counter;
 404
 405         evlist__for_each_entry(evsel_list, counter) {
 406                 if (counter->err)
 407                         pr_debug("failed to read counter %s\n", counter->name);
 408                 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
 409                         pr_warning("failed to process counter %s\n", counter->name);
 410                 counter->err = 0;
 411         }
 412
 413         perf_stat_merge_counters(&stat_config, evsel_list);
 414         perf_stat_process_percore(&stat_config, evsel_list);
 415 }
 416
 417 static void process_interval(void)
 418 {
 419         struct timespec ts, rs;
 420
 421         clock_gettime(CLOCK_MONOTONIC, &ts);
 422         diff_timespec(&rs, &ts, &ref_time);
 423
 424         evlist__reset_aggr_stats(evsel_list);
 425
 426         if (read_counters(&rs) == 0)
 427                 process_counters();
 428
 429         if (STAT_RECORD) {
 430                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 431                         pr_err("failed to write stat round event\n");
 432         }
 433
 434         init_stats(&walltime_nsecs_stats);
 435         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
 436         print_counters(&rs, 0, NULL);
 437 }
 438
 439 static bool handle_interval(unsigned int interval, int *times)
 440 {
 441         if (interval) {
 442                 process_interval();
 443                 if (interval_count && !(--(*times)))
 444                         return true;
 445         }
 446         return false;
 447 }
 448
 449 static int enable_counters(void)
 450 {
 451         struct evsel *evsel;
 452         int err;
 453
 454         evlist__for_each_entry(evsel_list, evsel) {
 455                 if (!evsel__is_bpf(evsel))
 456                         continue;
 457
 458                 err = bpf_counter__enable(evsel);
 459                 if (err)
 460                         return err;
 461         }
 462
 463         if (!target__enable_on_exec(&target)) {
 464                 if (!all_counters_use_bpf)
 465                         evlist__enable(evsel_list);
 466         }
 467         return 0;
 468 }
 469
 470 static void disable_counters(void)
 471 {
 472         struct evsel *counter;
 473
 474         /*
 475          * If we don't have tracee (attaching to task or cpu), counters may
 476          * still be running. To get accurate group ratios, we must stop groups
 477          * from counting before reading their constituent counters.
 478          */
 479         if (!target__none(&target)) {
 480                 evlist__for_each_entry(evsel_list, counter)
 481                         bpf_counter__disable(counter);
 482                 if (!all_counters_use_bpf)
 483                         evlist__disable(evsel_list);
 484         }
 485 }
 486
 487 static volatile sig_atomic_t workload_exec_errno;
 488
 489 /*
 490  * evlist__prepare_workload will send a SIGUSR1
 491  * if the fork fails, since we asked by setting its
 492  * want_signal to true.
 493  */
 494 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 495                                         void *ucontext __maybe_unused)
 496 {
 497         workload_exec_errno = info->si_value.sival_int;
 498 }
 499
 500 static bool evsel__should_store_id(struct evsel *counter)
 501 {
 502         return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
 503 }
 504
 505 static bool is_target_alive(struct target *_target,
 506                             struct perf_thread_map *threads)
 507 {
 508         struct stat st;
 509         int i;
 510
 511         if (!target__has_task(_target))
 512                 return true;
 513
 514         for (i = 0; i < threads->nr; i++) {
 515                 char path[PATH_MAX];
 516
 517                 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
 518                           threads->map[i].pid);
 519
 520                 if (!stat(path, &st))
 521                         return true;
 522         }
 523
 524         return false;
 525 }
 526
 527 static void process_evlist(struct evlist *evlist, unsigned int interval)
 528 {
 529         enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
 530
 531         if (evlist__ctlfd_process(evlist, &cmd) > 0) {
 532                 switch (cmd) {
 533                 case EVLIST_CTL_CMD_ENABLE:
 534                         fallthrough;
 535                 case EVLIST_CTL_CMD_DISABLE:
 536                         if (interval)
 537                                 process_interval();
 538                         break;
 539                 case EVLIST_CTL_CMD_SNAPSHOT:
 540                 case EVLIST_CTL_CMD_ACK:
 541                 case EVLIST_CTL_CMD_UNSUPPORTED:
 542                 case EVLIST_CTL_CMD_EVLIST:
 543                 case EVLIST_CTL_CMD_STOP:
 544                 case EVLIST_CTL_CMD_PING:
 545                 default:
 546                         break;
 547                 }
 548         }
 549 }
 550
 551 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
 552                         int *time_to_sleep)
 553 {
 554         int tts = *time_to_sleep;
 555         struct timespec time_diff;
 556
 557         diff_timespec(&time_diff, time_stop, time_start);
 558
 559         tts -= time_diff.tv_sec * MSEC_PER_SEC +
 560                time_diff.tv_nsec / NSEC_PER_MSEC;
 561
 562         if (tts < 0)
 563                 tts = 0;
 564
 565         *time_to_sleep = tts;
 566 }
 567
 568 static int dispatch_events(bool forks, int timeout, int interval, int *times)
 569 {
 570         int child_exited = 0, status = 0;
 571         int time_to_sleep, sleep_time;
 572         struct timespec time_start, time_stop;
 573
 574         if (interval)
 575                 sleep_time = interval;
 576         else if (timeout)
 577                 sleep_time = timeout;
 578         else
 579                 sleep_time = 1000;
 580
 581         time_to_sleep = sleep_time;
 582
 583         while (!done) {
 584                 if (forks)
 585                         child_exited = waitpid(child_pid, &status, WNOHANG);
 586                 else
 587                         child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
 588
 589                 if (child_exited)
 590                         break;
 591
 592                 clock_gettime(CLOCK_MONOTONIC, &time_start);
 593                 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
 594                         if (timeout || handle_interval(interval, times))
 595                                 break;
 596                         time_to_sleep = sleep_time;
 597                 } else { /* fd revent */
 598                         process_evlist(evsel_list, interval);
 599                         clock_gettime(CLOCK_MONOTONIC, &time_stop);
 600                         compute_tts(&time_start, &time_stop, &time_to_sleep);
 601                 }
 602         }
 603
 604         return status;
 605 }
 606
 607 enum counter_recovery {
 608         COUNTER_SKIP,
 609         COUNTER_RETRY,
 610         COUNTER_FATAL,
 611 };
 612
 613 static enum counter_recovery stat_handle_error(struct evsel *counter)
 614 {
 615         char msg[BUFSIZ];
 616         /*
 617          * PPC returns ENXIO for HW counters until 2.6.37
 618          * (behavior changed with commit b0a873e).
 619          */
 620         if (errno == EINVAL || errno == ENOSYS ||
 621             errno == ENOENT || errno == EOPNOTSUPP ||
 622             errno == ENXIO) {
 623                 if (verbose > 0)
 624                         ui__warning("%s event is not supported by the kernel.\n",
 625                                     evsel__name(counter));
 626                 counter->supported = false;
 627                 /*
 628                  * errored is a sticky flag that means one of the counter's
 629                  * cpu event had a problem and needs to be reexamined.
 630                  */
 631                 counter->errored = true;
 632
 633                 if ((evsel__leader(counter) != counter) ||
 634                     !(counter->core.leader->nr_members > 1))
 635                         return COUNTER_SKIP;
 636         } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
 637                 if (verbose > 0)
 638                         ui__warning("%s\n", msg);
 639                 return COUNTER_RETRY;
 640         } else if (target__has_per_thread(&target) &&
 641                    evsel_list->core.threads &&
 642                    evsel_list->core.threads->err_thread != -1) {
 643                 /*
 644                  * For global --per-thread case, skip current
 645                  * error thread.
 646                  */
 647                 if (!thread_map__remove(evsel_list->core.threads,
 648                                         evsel_list->core.threads->err_thread)) {
 649                         evsel_list->core.threads->err_thread = -1;
 650                         return COUNTER_RETRY;
 651                 }
 652         } else if (counter->skippable) {
 653                 if (verbose > 0)
 654                         ui__warning("skipping event %s that kernel failed to open .\n",
 655                                     evsel__name(counter));
 656                 counter->supported = false;
 657                 counter->errored = true;
 658                 return COUNTER_SKIP;
 659         }
 660
 661         evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
 662         ui__error("%s\n", msg);
 663
 664         if (child_pid != -1)
 665                 kill(child_pid, SIGTERM);
 666         return COUNTER_FATAL;
 667 }
 668
 669 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 670 {
 671         int interval = stat_config.interval;
 672         int times = stat_config.times;
 673         int timeout = stat_config.timeout;
 674         char msg[BUFSIZ];
 675         unsigned long long t0, t1;
 676         struct evsel *counter;
 677         size_t l;
 678         int status = 0;
 679         const bool forks = (argc > 0);
 680         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 681         struct evlist_cpu_iterator evlist_cpu_itr;
 682         struct affinity saved_affinity, *affinity = NULL;
 683         int err;
 684         bool second_pass = false;
 685
 686         if (forks) {
 687                 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
 688                         perror("failed to prepare workload");
 689                         return -1;
 690                 }
 691                 child_pid = evsel_list->workload.pid;
 692         }
 693
 694         if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
 695                 if (affinity__setup(&saved_affinity) < 0)
 696                         return -1;
 697                 affinity = &saved_affinity;
 698         }
 699
 700         evlist__for_each_entry(evsel_list, counter) {
 701                 counter->reset_group = false;
 702                 if (bpf_counter__load(counter, &target))
 703                         return -1;
 704                 if (!(evsel__is_bperf(counter)))
 705                         all_counters_use_bpf = false;
 706         }
 707
 708         evlist__reset_aggr_stats(evsel_list);
 709
 710         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 711                 counter = evlist_cpu_itr.evsel;
 712
 713                 /*
 714                  * bperf calls evsel__open_per_cpu() in bperf__load(), so
 715                  * no need to call it again here.
 716                  */
 717                 if (target.use_bpf)
 718                         break;
 719
 720                 if (counter->reset_group || counter->errored)
 721                         continue;
 722                 if (evsel__is_bperf(counter))
 723                         continue;
 724 try_again:
 725                 if (create_perf_stat_counter(counter, &stat_config, &target,
 726                                              evlist_cpu_itr.cpu_map_idx) < 0) {
 727
 728                         /*
 729                          * Weak group failed. We cannot just undo this here
 730                          * because earlier CPUs might be in group mode, and the kernel
 731                          * doesn't support mixing group and non group reads. Defer
 732                          * it to later.
 733                          * Don't close here because we're in the wrong affinity.
 734                          */
 735                         if ((errno == EINVAL || errno == EBADF) &&
 736                                 evsel__leader(counter) != counter &&
 737                                 counter->weak_group) {
 738                                 evlist__reset_weak_group(evsel_list, counter, false);
 739                                 assert(counter->reset_group);
 740                                 second_pass = true;
 741                                 continue;
 742                         }
 743
 744                         switch (stat_handle_error(counter)) {
 745                         case COUNTER_FATAL:
 746                                 return -1;
 747                         case COUNTER_RETRY:
 748                                 goto try_again;
 749                         case COUNTER_SKIP:
 750                                 continue;
 751                         default:
 752                                 break;
 753                         }
 754
 755                 }
 756                 counter->supported = true;
 757         }
 758
 759         if (second_pass) {
 760                 /*
 761                  * Now redo all the weak group after closing them,
 762                  * and also close errored counters.
 763                  */
 764
 765                 /* First close errored or weak retry */
 766                 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 767                         counter = evlist_cpu_itr.evsel;
 768
 769                         if (!counter->reset_group && !counter->errored)
 770                                 continue;
 771
 772                         perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
 773                 }
 774                 /* Now reopen weak */
 775                 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 776                         counter = evlist_cpu_itr.evsel;
 777
 778                         if (!counter->reset_group)
 779                                 continue;
 780 try_again_reset:
 781                         pr_debug2("reopening weak %s\n", evsel__name(counter));
 782                         if (create_perf_stat_counter(counter, &stat_config, &target,
 783                                                      evlist_cpu_itr.cpu_map_idx) < 0) {
 784
 785                                 switch (stat_handle_error(counter)) {
 786                                 case COUNTER_FATAL:
 787                                         return -1;
 788                                 case COUNTER_RETRY:
 789                                         goto try_again_reset;
 790                                 case COUNTER_SKIP:
 791                                         continue;
 792                                 default:
 793                                         break;
 794                                 }
 795                         }
 796                         counter->supported = true;
 797                 }
 798         }
 799         affinity__cleanup(affinity);
 800
 801         evlist__for_each_entry(evsel_list, counter) {
 802                 if (!counter->supported) {
 803                         perf_evsel__free_fd(&counter->core);
 804                         continue;
 805                 }
 806
 807                 l = strlen(counter->unit);
 808                 if (l > stat_config.unit_width)
 809                         stat_config.unit_width = l;
 810
 811                 if (evsel__should_store_id(counter) &&
 812                     evsel__store_ids(counter, evsel_list))
 813                         return -1;
 814         }
 815
 816         if (evlist__apply_filters(evsel_list, &counter)) {
 817                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 818                         counter->filter, evsel__name(counter), errno,
 819                         str_error_r(errno, msg, sizeof(msg)));
 820                 return -1;
 821         }
 822
 823         if (STAT_RECORD) {
 824                 int fd = perf_data__fd(&perf_stat.data);
 825
 826                 if (is_pipe) {
 827                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 828                 } else {
 829                         err = perf_session__write_header(perf_stat.session, evsel_list,
 830                                                          fd, false);
 831                 }
 832
 833                 if (err < 0)
 834                         return err;
 835
 836                 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
 837                                                          process_synthesized_event, is_pipe);
 838                 if (err < 0)
 839                         return err;
 840         }
 841
 842         if (target.initial_delay) {
 843                 pr_info(EVLIST_DISABLED_MSG);
 844         } else {
 845                 err = enable_counters();
 846                 if (err)
 847                         return -1;
 848         }
 849
 850         /* Exec the command, if any */
 851         if (forks)
 852                 evlist__start_workload(evsel_list);
 853
 854         if (target.initial_delay > 0) {
 855                 usleep(target.initial_delay * USEC_PER_MSEC);
 856                 err = enable_counters();
 857                 if (err)
 858                         return -1;
 859
 860                 pr_info(EVLIST_ENABLED_MSG);
 861         }
 862
 863         t0 = rdclock();
 864         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 865
 866         if (forks) {
 867                 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
 868                         status = dispatch_events(forks, timeout, interval, &times);
 869                 if (child_pid != -1) {
 870                         if (timeout)
 871                                 kill(child_pid, SIGTERM);
 872                         wait4(child_pid, &status, 0, &stat_config.ru_data);
 873                 }
 874
 875                 if (workload_exec_errno) {
 876                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 877                         pr_err("Workload failed: %s\n", emsg);
 878                         return -1;
 879                 }
 880
 881                 if (WIFSIGNALED(status))
 882                         psignal(WTERMSIG(status), argv[0]);
 883         } else {
 884                 status = dispatch_events(forks, timeout, interval, &times);
 885         }
 886
 887         disable_counters();
 888
 889         t1 = rdclock();
 890
 891         if (stat_config.walltime_run_table)
 892                 stat_config.walltime_run[run_idx] = t1 - t0;
 893
 894         if (interval && stat_config.summary) {
 895                 stat_config.interval = 0;
 896                 stat_config.stop_read_counter = true;
 897                 init_stats(&walltime_nsecs_stats);
 898                 update_stats(&walltime_nsecs_stats, t1 - t0);
 899
 900                 evlist__copy_prev_raw_counts(evsel_list);
 901                 evlist__reset_prev_raw_counts(evsel_list);
 902                 evlist__reset_aggr_stats(evsel_list);
 903         } else {
 904                 update_stats(&walltime_nsecs_stats, t1 - t0);
 905                 update_rusage_stats(&ru_stats, &stat_config.ru_data);
 906         }
 907
 908         /*
 909          * Closing a group leader splits the group, and as we only disable
 910          * group leaders, results in remaining events becoming enabled. To
 911          * avoid arbitrary skew, we must read all counters before closing any
 912          * group leaders.
 913          */
 914         if (read_counters(&(struct timespec) { .tv_nsec = t1-t0 }) == 0)
 915                 process_counters();
 916
 917         /*
 918          * We need to keep evsel_list alive, because it's processed
 919          * later the evsel_list will be closed after.
 920          */
 921         if (!STAT_RECORD)
 922                 evlist__close(evsel_list);
 923
 924         return WEXITSTATUS(status);
 925 }
 926
 927 static int run_perf_stat(int argc, const char **argv, int run_idx)
 928 {
 929         int ret;
 930
 931         if (pre_cmd) {
 932                 ret = system(pre_cmd);
 933                 if (ret)
 934                         return ret;
 935         }
 936
 937         if (sync_run)
 938                 sync();
 939
 940         ret = __run_perf_stat(argc, argv, run_idx);
 941         if (ret)
 942                 return ret;
 943
 944         if (post_cmd) {
 945                 ret = system(post_cmd);
 946                 if (ret)
 947                         return ret;
 948         }
 949
 950         return ret;
 951 }
 952
 953 static void print_counters(struct timespec *ts, int argc, const char **argv)
 954 {
 955         /* Do not print anything if we record to the pipe. */
 956         if (STAT_RECORD && perf_stat.data.is_pipe)
 957                 return;
 958         if (quiet)
 959                 return;
 960
 961         evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
 962 }
 963
 964 static volatile sig_atomic_t signr = -1;
 965
 966 static void skip_signal(int signo)
 967 {
 968         if ((child_pid == -1) || stat_config.interval)
 969                 done = 1;
 970
 971         signr = signo;
 972         /*
 973          * render child_pid harmless
 974          * won't send SIGTERM to a random
 975          * process in case of race condition
 976          * and fast PID recycling
 977          */
 978         child_pid = -1;
 979 }
 980
 981 static void sig_atexit(void)
 982 {
 983         sigset_t set, oset;
 984
 985         /*
 986          * avoid race condition with SIGCHLD handler
 987          * in skip_signal() which is modifying child_pid
 988          * goal is to avoid send SIGTERM to a random
 989          * process
 990          */
 991         sigemptyset(&set);
 992         sigaddset(&set, SIGCHLD);
 993         sigprocmask(SIG_BLOCK, &set, &oset);
 994
 995         if (child_pid != -1)
 996                 kill(child_pid, SIGTERM);
 997
 998         sigprocmask(SIG_SETMASK, &oset, NULL);
 999
1000         if (signr == -1)
1001                 return;
1002
1003         signal(signr, SIG_DFL);
1004         kill(getpid(), signr);
1005 }
1006
1007 void perf_stat__set_big_num(int set)
1008 {
1009         stat_config.big_num = (set != 0);
1010 }
1011
1012 void perf_stat__set_no_csv_summary(int set)
1013 {
1014         stat_config.no_csv_summary = (set != 0);
1015 }
1016
1017 static int stat__set_big_num(const struct option *opt __maybe_unused,
1018                              const char *s __maybe_unused, int unset)
1019 {
1020         big_num_opt = unset ? 0 : 1;
1021         perf_stat__set_big_num(!unset);
1022         return 0;
1023 }
1024
1025 static int enable_metric_only(const struct option *opt __maybe_unused,
1026                               const char *s __maybe_unused, int unset)
1027 {
1028         force_metric_only = true;
1029         stat_config.metric_only = !unset;
1030         return 0;
1031 }
1032
1033 static int append_metric_groups(const struct option *opt __maybe_unused,
1034                                const char *str,
1035                                int unset __maybe_unused)
1036 {
1037         if (metrics) {
1038                 char *tmp;
1039
1040                 if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1041                         return -ENOMEM;
1042                 free(metrics);
1043                 metrics = tmp;
1044         } else {
1045                 metrics = strdup(str);
1046                 if (!metrics)
1047                         return -ENOMEM;
1048         }
1049         return 0;
1050 }
1051
1052 static int parse_control_option(const struct option *opt,
1053                                 const char *str,
1054                                 int unset __maybe_unused)
1055 {
1056         struct perf_stat_config *config = opt->value;
1057
1058         return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1059 }
1060
1061 static int parse_stat_cgroups(const struct option *opt,
1062                               const char *str, int unset)
1063 {
1064         if (stat_config.cgroup_list) {
1065                 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1066                 return -1;
1067         }
1068
1069         return parse_cgroups(opt, str, unset);
1070 }
1071
1072 static int parse_cputype(const struct option *opt,
1073                              const char *str,
1074                              int unset __maybe_unused)
1075 {
1076         const struct perf_pmu *pmu;
1077         struct evlist *evlist = *(struct evlist **)opt->value;
1078
1079         if (!list_empty(&evlist->core.entries)) {
1080                 fprintf(stderr, "Must define cputype before events/metrics\n");
1081                 return -1;
1082         }
1083
1084         pmu = perf_pmus__pmu_for_pmu_filter(str);
1085         if (!pmu) {
1086                 fprintf(stderr, "--cputype %s is not supported!\n", str);
1087                 return -1;
1088         }
1089         parse_events_option_args.pmu_filter = pmu->name;
1090
1091         return 0;
1092 }
1093
1094 static int parse_cache_level(const struct option *opt,
1095                              const char *str,
1096                              int unset __maybe_unused)
1097 {
1098         int level;
1099         u32 *aggr_mode = (u32 *)opt->value;
1100         u32 *aggr_level = (u32 *)opt->data;
1101
1102         /*
1103          * If no string is specified, aggregate based on the topology of
1104          * Last Level Cache (LLC). Since the LLC level can change from
1105          * architecture to architecture, set level greater than
1106          * MAX_CACHE_LVL which will be interpreted as LLC.
1107          */
1108         if (str == NULL) {
1109                 level = MAX_CACHE_LVL + 1;
1110                 goto out;
1111         }
1112
1113         /*
1114          * The format to specify cache level is LX or lX where X is the
1115          * cache level.
1116          */
1117         if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
1118                 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1119                        MAX_CACHE_LVL,
1120                        MAX_CACHE_LVL);
1121                 return -EINVAL;
1122         }
1123
1124         level = atoi(&str[1]);
1125         if (level < 1) {
1126                 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1127                        MAX_CACHE_LVL,
1128                        MAX_CACHE_LVL);
1129                 return -EINVAL;
1130         }
1131
1132         if (level > MAX_CACHE_LVL) {
1133                 pr_err("perf only supports max cache level of %d.\n"
1134                        "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
1135                 return -EINVAL;
1136         }
1137 out:
1138         *aggr_mode = AGGR_CACHE;
1139         *aggr_level = level;
1140         return 0;
1141 }
1142
1143 static struct option stat_options[] = {
1144         OPT_BOOLEAN('T', "transaction", &transaction_run,
1145                     "hardware transaction statistics"),
1146         OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
1147                      "event selector. use 'perf list' to list available events",
1148                      parse_events_option),
1149         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1150                      "event filter", parse_filter),
1151         OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1152                     "child tasks do not inherit counters"),
1153         OPT_STRING('p', "pid", &target.pid, "pid",
1154                    "stat events on existing process id"),
1155         OPT_STRING('t', "tid", &target.tid, "tid",
1156                    "stat events on existing thread id"),
1157 #ifdef HAVE_BPF_SKEL
1158         OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1159                    "stat events on existing bpf program id"),
1160         OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1161                     "use bpf program to count events"),
1162         OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1163                    "path to perf_event_attr map"),
1164 #endif
1165         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1166                     "system-wide collection from all CPUs"),
1167         OPT_BOOLEAN(0, "scale", &stat_config.scale,
1168                     "Use --no-scale to disable counter scaling for multiplexing"),
1169         OPT_INCR('v', "verbose", &verbose,
1170                     "be more verbose (show counter open errors, etc)"),
1171         OPT_INTEGER('r', "repeat", &stat_config.run_count,
1172                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1173         OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1174                     "display details about each run (only with -r option)"),
1175         OPT_BOOLEAN('n', "null", &stat_config.null_run,
1176                     "null run - dont start any counters"),
1177         OPT_INCR('d', "detailed", &detailed_run,
1178                     "detailed run - start a lot of events"),
1179         OPT_BOOLEAN('S', "sync", &sync_run,
1180                     "call sync() before starting a run"),
1181         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1182                            "print large numbers with thousands\' separators",
1183                            stat__set_big_num),
1184         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1185                     "list of cpus to monitor in system-wide"),
1186         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1187                     "disable aggregation across CPUs or PMUs", AGGR_NONE),
1188         OPT_SET_UINT(0, "no-merge", &stat_config.aggr_mode,
1189                     "disable aggregation the same as -A or -no-aggr", AGGR_NONE),
1190         OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1191                     "Merge identical named hybrid events"),
1192         OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1193                    "print counts with custom separator"),
1194         OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1195                    "print counts in JSON format"),
1196         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1197                      "monitor event in cgroup name only", parse_stat_cgroups),
1198         OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1199                     "expand events for each cgroup"),
1200         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1201         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1202         OPT_INTEGER(0, "log-fd", &output_fd,
1203                     "log output to fd, instead of stderr"),
1204         OPT_STRING(0, "pre", &pre_cmd, "command",
1205                         "command to run prior to the measured command"),
1206         OPT_STRING(0, "post", &post_cmd, "command",
1207                         "command to run after to the measured command"),
1208         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1209                     "print counts at regular interval in ms "
1210                     "(overhead is possible for values <= 100ms)"),
1211         OPT_INTEGER(0, "interval-count", &stat_config.times,
1212                     "print counts for fixed number of times"),
1213         OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1214                     "clear screen in between new interval"),
1215         OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1216                     "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1217         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1218                      "aggregate counts per processor socket", AGGR_SOCKET),
1219         OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1220                      "aggregate counts per processor die", AGGR_DIE),
1221         OPT_SET_UINT(0, "per-cluster", &stat_config.aggr_mode,
1222                      "aggregate counts per processor cluster", AGGR_CLUSTER),
1223         OPT_CALLBACK_OPTARG(0, "per-cache", &stat_config.aggr_mode, &stat_config.aggr_level,
1224                             "cache level", "aggregate count at this cache level (Default: LLC)",
1225                             parse_cache_level),
1226         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1227                      "aggregate counts per physical processor core", AGGR_CORE),
1228         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1229                      "aggregate counts per thread", AGGR_THREAD),
1230         OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1231                      "aggregate counts per numa node", AGGR_NODE),
1232         OPT_INTEGER('D', "delay", &target.initial_delay,
1233                     "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1234         OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1235                         "Only print computed metrics. No raw values", enable_metric_only),
1236         OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1237                        "don't group metric events, impacts multiplexing"),
1238         OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1239                        "don't try to share events between metrics in a group"),
1240         OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold,
1241                        "disable adding events for the metric threshold calculation"),
1242         OPT_BOOLEAN(0, "topdown", &topdown_run,
1243                         "measure top-down statistics"),
1244         OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1245                         "Set the metrics level for the top-down statistics (0: max level)"),
1246         OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1247                         "measure SMI cost"),
1248         OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1249                      "monitor specified metrics or metric groups (separated by ,)",
1250                      append_metric_groups),
1251         OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1252                          "Configure all used events to run in kernel space.",
1253                          PARSE_OPT_EXCLUSIVE),
1254         OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1255                          "Configure all used events to run in user space.",
1256                          PARSE_OPT_EXCLUSIVE),
1257         OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1258                     "Use with 'percore' event qualifier to show the event "
1259                     "counts of one hardware thread by sum up total hardware "
1260                     "threads of same physical core"),
1261         OPT_BOOLEAN(0, "summary", &stat_config.summary,
1262                        "print summary for interval mode"),
1263         OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1264                        "don't print 'summary' for CSV summary output"),
1265         OPT_BOOLEAN(0, "quiet", &quiet,
1266                         "don't print any output, messages or warnings (useful with record)"),
1267         OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1268                      "Only enable events on applying cpu with this type "
1269                      "for hybrid platform (e.g. core or atom)",
1270                      parse_cputype),
1271 #ifdef HAVE_LIBPFM
1272         OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1273                 "libpfm4 event selector. use 'perf list' to list available events",
1274                 parse_libpfm_events_option),
1275 #endif
1276         OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1277                      "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1278                      "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1279                      "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1280                       parse_control_option),
1281         OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1282                             "measure I/O performance metrics provided by arch/platform",
1283                             iostat_parse),
1284         OPT_END()
1285 };
1286
1287 /**
1288  * Calculate the cache instance ID from the map in
1289  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1290  * Cache instance ID is the first CPU reported in the shared_cpu_list file.
1291  */
1292 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
1293 {
1294         int id;
1295         struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
1296
1297         /*
1298          * If the map contains no CPU, consider the current CPU to
1299          * be the first online CPU in the cache domain else use the
1300          * first online CPU of the cache domain as the ID.
1301          */
1302         id = perf_cpu_map__min(cpu_map).cpu;
1303         if (id == -1)
1304                 id = cpu.cpu;
1305
1306         /* Free the perf_cpu_map used to find the cache ID */
1307         perf_cpu_map__put(cpu_map);
1308
1309         return id;
1310 }
1311
1312 /**
1313  * cpu__get_cache_id - Returns 0 if successful in populating the
1314  * cache level and cache id. Cache level is read from
1315  * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
1316  * is the first CPU reported by
1317  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1318  */
1319 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
1320 {
1321         int ret = 0;
1322         u32 cache_level = stat_config.aggr_level;
1323         struct cpu_cache_level caches[MAX_CACHE_LVL];
1324         u32 i = 0, caches_cnt = 0;
1325
1326         cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1327         cache->cache = -1;
1328
1329         ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
1330         if (ret) {
1331                 /*
1332                  * If caches_cnt is not 0, cpu_cache_level data
1333                  * was allocated when building the topology.
1334                  * Free the allocated data before returning.
1335                  */
1336                 if (caches_cnt)
1337                         goto free_caches;
1338
1339                 return ret;
1340         }
1341
1342         if (!caches_cnt)
1343                 return -1;
1344
1345         /*
1346          * Save the data for the highest level if no
1347          * level was specified by the user.
1348          */
1349         if (cache_level > MAX_CACHE_LVL) {
1350                 int max_level_index = 0;
1351
1352                 for (i = 1; i < caches_cnt; ++i) {
1353                         if (caches[i].level > caches[max_level_index].level)
1354                                 max_level_index = i;
1355                 }
1356
1357                 cache->cache_lvl = caches[max_level_index].level;
1358                 cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
1359
1360                 /* Reset i to 0 to free entire caches[] */
1361                 i = 0;
1362                 goto free_caches;
1363         }
1364
1365         for (i = 0; i < caches_cnt; ++i) {
1366                 if (caches[i].level == cache_level) {
1367                         cache->cache_lvl = cache_level;
1368                         cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1369                 }
1370
1371                 cpu_cache_level__free(&caches[i]);
1372         }
1373
1374 free_caches:
1375         /*
1376          * Free all the allocated cpu_cache_level data.
1377          */
1378         while (i < caches_cnt)
1379                 cpu_cache_level__free(&caches[i++]);
1380
1381         return ret;
1382 }
1383
1384 /**
1385  * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
1386  * level, die and socket populated with the cache instache ID, cache level,
1387  * die and socket for cpu. The function signature is compatible with
1388  * aggr_cpu_id_get_t.
1389  */
1390 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
1391 {
1392         int ret;
1393         struct aggr_cpu_id id;
1394         struct perf_cache cache;
1395
1396         id = aggr_cpu_id__die(cpu, data);
1397         if (aggr_cpu_id__is_empty(&id))
1398                 return id;
1399
1400         ret = cpu__get_cache_details(cpu, &cache);
1401         if (ret)
1402                 return id;
1403
1404         id.cache_lvl = cache.cache_lvl;
1405         id.cache = cache.cache;
1406         return id;
1407 }
1408
1409 static const char *const aggr_mode__string[] = {
1410         [AGGR_CORE] = "core",
1411         [AGGR_CACHE] = "cache",
1412         [AGGR_CLUSTER] = "cluster",
1413         [AGGR_DIE] = "die",
1414         [AGGR_GLOBAL] = "global",
1415         [AGGR_NODE] = "node",
1416         [AGGR_NONE] = "none",
1417         [AGGR_SOCKET] = "socket",
1418         [AGGR_THREAD] = "thread",
1419         [AGGR_UNSET] = "unset",
1420 };
1421
1422 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1423                                                 struct perf_cpu cpu)
1424 {
1425         return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1426 }
1427
1428 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1429                                              struct perf_cpu cpu)
1430 {
1431         return aggr_cpu_id__die(cpu, /*data=*/NULL);
1432 }
1433
1434 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
1435                                                   struct perf_cpu cpu)
1436 {
1437         return aggr_cpu_id__cache(cpu, /*data=*/NULL);
1438 }
1439
1440 static struct aggr_cpu_id perf_stat__get_cluster(struct perf_stat_config *config __maybe_unused,
1441                                                  struct perf_cpu cpu)
1442 {
1443         return aggr_cpu_id__cluster(cpu, /*data=*/NULL);
1444 }
1445
1446 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1447                                               struct perf_cpu cpu)
1448 {
1449         return aggr_cpu_id__core(cpu, /*data=*/NULL);
1450 }
1451
1452 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1453                                               struct perf_cpu cpu)
1454 {
1455         return aggr_cpu_id__node(cpu, /*data=*/NULL);
1456 }
1457
1458 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1459                                                 struct perf_cpu cpu)
1460 {
1461         return aggr_cpu_id__global(cpu, /*data=*/NULL);
1462 }
1463
1464 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1465                                              struct perf_cpu cpu)
1466 {
1467         return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1468 }
1469
1470 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1471                                               aggr_get_id_t get_id, struct perf_cpu cpu)
1472 {
1473         struct aggr_cpu_id id;
1474
1475         /* per-process mode - should use global aggr mode */
1476         if (cpu.cpu == -1)
1477                 return get_id(config, cpu);
1478
1479         if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1480                 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1481
1482         id = config->cpus_aggr_map->map[cpu.cpu];
1483         return id;
1484 }
1485
1486 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1487                                                        struct perf_cpu cpu)
1488 {
1489         return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1490 }
1491
1492 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1493                                                     struct perf_cpu cpu)
1494 {
1495         return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1496 }
1497
1498 static struct aggr_cpu_id perf_stat__get_cluster_cached(struct perf_stat_config *config,
1499                                                         struct perf_cpu cpu)
1500 {
1501         return perf_stat__get_aggr(config, perf_stat__get_cluster, cpu);
1502 }
1503
1504 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
1505                                                          struct perf_cpu cpu)
1506 {
1507         return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
1508 }
1509
1510 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1511                                                      struct perf_cpu cpu)
1512 {
1513         return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1514 }
1515
1516 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1517                                                      struct perf_cpu cpu)
1518 {
1519         return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1520 }
1521
1522 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1523                                                        struct perf_cpu cpu)
1524 {
1525         return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1526 }
1527
1528 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1529                                                     struct perf_cpu cpu)
1530 {
1531         return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1532 }
1533
1534 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1535 {
1536         switch (aggr_mode) {
1537         case AGGR_SOCKET:
1538                 return aggr_cpu_id__socket;
1539         case AGGR_DIE:
1540                 return aggr_cpu_id__die;
1541         case AGGR_CLUSTER:
1542                 return aggr_cpu_id__cluster;
1543         case AGGR_CACHE:
1544                 return aggr_cpu_id__cache;
1545         case AGGR_CORE:
1546                 return aggr_cpu_id__core;
1547         case AGGR_NODE:
1548                 return aggr_cpu_id__node;
1549         case AGGR_NONE:
1550                 return aggr_cpu_id__cpu;
1551         case AGGR_GLOBAL:
1552                 return aggr_cpu_id__global;
1553         case AGGR_THREAD:
1554         case AGGR_UNSET:
1555         case AGGR_MAX:
1556         default:
1557                 return NULL;
1558         }
1559 }
1560
1561 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1562 {
1563         switch (aggr_mode) {
1564         case AGGR_SOCKET:
1565                 return perf_stat__get_socket_cached;
1566         case AGGR_DIE:
1567                 return perf_stat__get_die_cached;
1568         case AGGR_CLUSTER:
1569                 return perf_stat__get_cluster_cached;
1570         case AGGR_CACHE:
1571                 return perf_stat__get_cache_id_cached;
1572         case AGGR_CORE:
1573                 return perf_stat__get_core_cached;
1574         case AGGR_NODE:
1575                 return perf_stat__get_node_cached;
1576         case AGGR_NONE:
1577                 return perf_stat__get_cpu_cached;
1578         case AGGR_GLOBAL:
1579                 return perf_stat__get_global_cached;
1580         case AGGR_THREAD:
1581         case AGGR_UNSET:
1582         case AGGR_MAX:
1583         default:
1584                 return NULL;
1585         }
1586 }
1587
1588 static int perf_stat_init_aggr_mode(void)
1589 {
1590         int nr;
1591         aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1592
1593         if (get_id) {
1594                 bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1595                 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1596                                                          get_id, /*data=*/NULL, needs_sort);
1597                 if (!stat_config.aggr_map) {
1598                         pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1599                         return -1;
1600                 }
1601                 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1602         }
1603
1604         if (stat_config.aggr_mode == AGGR_THREAD) {
1605                 nr = perf_thread_map__nr(evsel_list->core.threads);
1606                 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1607                 if (stat_config.aggr_map == NULL)
1608                         return -ENOMEM;
1609
1610                 for (int s = 0; s < nr; s++) {
1611                         struct aggr_cpu_id id = aggr_cpu_id__empty();
1612
1613                         id.thread_idx = s;
1614                         stat_config.aggr_map->map[s] = id;
1615                 }
1616                 return 0;
1617         }
1618
1619         /*
1620          * The evsel_list->cpus is the base we operate on,
1621          * taking the highest cpu number to be the size of
1622          * the aggregation translate cpumap.
1623          */
1624         if (!perf_cpu_map__is_any_cpu_or_is_empty(evsel_list->core.user_requested_cpus))
1625                 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1626         else
1627                 nr = 0;
1628         stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1629         return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1630 }
1631
1632 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1633 {
1634         free(map);
1635 }
1636
1637 static void perf_stat__exit_aggr_mode(void)
1638 {
1639         cpu_aggr_map__delete(stat_config.aggr_map);
1640         cpu_aggr_map__delete(stat_config.cpus_aggr_map);
1641         stat_config.aggr_map = NULL;
1642         stat_config.cpus_aggr_map = NULL;
1643 }
1644
1645 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1646 {
1647         struct perf_env *env = data;
1648         struct aggr_cpu_id id = aggr_cpu_id__empty();
1649
1650         if (cpu.cpu != -1)
1651                 id.socket = env->cpu[cpu.cpu].socket_id;
1652
1653         return id;
1654 }
1655
1656 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1657 {
1658         struct perf_env *env = data;
1659         struct aggr_cpu_id id = aggr_cpu_id__empty();
1660
1661         if (cpu.cpu != -1) {
1662                 /*
1663                  * die_id is relative to socket, so start
1664                  * with the socket ID and then add die to
1665                  * make a unique ID.
1666                  */
1667                 id.socket = env->cpu[cpu.cpu].socket_id;
1668                 id.die = env->cpu[cpu.cpu].die_id;
1669         }
1670
1671         return id;
1672 }
1673
1674 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
1675                                            u32 cache_level, struct aggr_cpu_id *id)
1676 {
1677         int i;
1678         int caches_cnt = env->caches_cnt;
1679         struct cpu_cache_level *caches = env->caches;
1680
1681         id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1682         id->cache = -1;
1683
1684         if (!caches_cnt)
1685                 return;
1686
1687         for (i = caches_cnt - 1; i > -1; --i) {
1688                 struct perf_cpu_map *cpu_map;
1689                 int map_contains_cpu;
1690
1691                 /*
1692                  * If user has not specified a level, find the fist level with
1693                  * the cpu in the map. Since building the map is expensive, do
1694                  * this only if levels match.
1695                  */
1696                 if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
1697                         continue;
1698
1699                 cpu_map = perf_cpu_map__new(caches[i].map);
1700                 map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
1701                 perf_cpu_map__put(cpu_map);
1702
1703                 if (map_contains_cpu != -1) {
1704                         id->cache_lvl = caches[i].level;
1705                         id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1706                         return;
1707                 }
1708         }
1709 }
1710
1711 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
1712                                                           void *data)
1713 {
1714         struct perf_env *env = data;
1715         struct aggr_cpu_id id = aggr_cpu_id__empty();
1716
1717         if (cpu.cpu != -1) {
1718                 u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
1719
1720                 id.socket = env->cpu[cpu.cpu].socket_id;
1721                 id.die = env->cpu[cpu.cpu].die_id;
1722                 perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
1723         }
1724
1725         return id;
1726 }
1727
1728 static struct aggr_cpu_id perf_env__get_cluster_aggr_by_cpu(struct perf_cpu cpu,
1729                                                             void *data)
1730 {
1731         struct perf_env *env = data;
1732         struct aggr_cpu_id id = aggr_cpu_id__empty();
1733
1734         if (cpu.cpu != -1) {
1735                 id.socket = env->cpu[cpu.cpu].socket_id;
1736                 id.die = env->cpu[cpu.cpu].die_id;
1737                 id.cluster = env->cpu[cpu.cpu].cluster_id;
1738         }
1739
1740         return id;
1741 }
1742
1743 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1744 {
1745         struct perf_env *env = data;
1746         struct aggr_cpu_id id = aggr_cpu_id__empty();
1747
1748         if (cpu.cpu != -1) {
1749                 /*
1750                  * core_id is relative to socket, die and cluster, we need a
1751                  * global id. So we set socket, die id, cluster id and core id.
1752                  */
1753                 id.socket = env->cpu[cpu.cpu].socket_id;
1754                 id.die = env->cpu[cpu.cpu].die_id;
1755                 id.cluster = env->cpu[cpu.cpu].cluster_id;
1756                 id.core = env->cpu[cpu.cpu].core_id;
1757         }
1758
1759         return id;
1760 }
1761
1762 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1763 {
1764         struct perf_env *env = data;
1765         struct aggr_cpu_id id = aggr_cpu_id__empty();
1766
1767         if (cpu.cpu != -1) {
1768                 /*
1769                  * core_id is relative to socket and die,
1770                  * we need a global id. So we set
1771                  * socket, die id and core id
1772                  */
1773                 id.socket = env->cpu[cpu.cpu].socket_id;
1774                 id.die = env->cpu[cpu.cpu].die_id;
1775                 id.core = env->cpu[cpu.cpu].core_id;
1776                 id.cpu = cpu;
1777         }
1778
1779         return id;
1780 }
1781
1782 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1783 {
1784         struct aggr_cpu_id id = aggr_cpu_id__empty();
1785
1786         id.node = perf_env__numa_node(data, cpu);
1787         return id;
1788 }
1789
1790 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1791                                                            void *data __maybe_unused)
1792 {
1793         struct aggr_cpu_id id = aggr_cpu_id__empty();
1794
1795         /* it always aggregates to the cpu 0 */
1796         id.cpu = (struct perf_cpu){ .cpu = 0 };
1797         return id;
1798 }
1799
1800 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1801                                                      struct perf_cpu cpu)
1802 {
1803         return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1804 }
1805 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1806                                                   struct perf_cpu cpu)
1807 {
1808         return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1809 }
1810
1811 static struct aggr_cpu_id perf_stat__get_cluster_file(struct perf_stat_config *config __maybe_unused,
1812                                                       struct perf_cpu cpu)
1813 {
1814         return perf_env__get_cluster_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1815 }
1816
1817 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
1818                                                     struct perf_cpu cpu)
1819 {
1820         return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1821 }
1822
1823 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1824                                                    struct perf_cpu cpu)
1825 {
1826         return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1827 }
1828
1829 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1830                                                   struct perf_cpu cpu)
1831 {
1832         return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1833 }
1834
1835 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1836                                                    struct perf_cpu cpu)
1837 {
1838         return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1839 }
1840
1841 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1842                                                      struct perf_cpu cpu)
1843 {
1844         return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1845 }
1846
1847 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1848 {
1849         switch (aggr_mode) {
1850         case AGGR_SOCKET:
1851                 return perf_env__get_socket_aggr_by_cpu;
1852         case AGGR_DIE:
1853                 return perf_env__get_die_aggr_by_cpu;
1854         case AGGR_CLUSTER:
1855                 return perf_env__get_cluster_aggr_by_cpu;
1856         case AGGR_CACHE:
1857                 return perf_env__get_cache_aggr_by_cpu;
1858         case AGGR_CORE:
1859                 return perf_env__get_core_aggr_by_cpu;
1860         case AGGR_NODE:
1861                 return perf_env__get_node_aggr_by_cpu;
1862         case AGGR_GLOBAL:
1863                 return perf_env__get_global_aggr_by_cpu;
1864         case AGGR_NONE:
1865                 return perf_env__get_cpu_aggr_by_cpu;
1866         case AGGR_THREAD:
1867         case AGGR_UNSET:
1868         case AGGR_MAX:
1869         default:
1870                 return NULL;
1871         }
1872 }
1873
1874 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1875 {
1876         switch (aggr_mode) {
1877         case AGGR_SOCKET:
1878                 return perf_stat__get_socket_file;
1879         case AGGR_DIE:
1880                 return perf_stat__get_die_file;
1881         case AGGR_CLUSTER:
1882                 return perf_stat__get_cluster_file;
1883         case AGGR_CACHE:
1884                 return perf_stat__get_cache_file;
1885         case AGGR_CORE:
1886                 return perf_stat__get_core_file;
1887         case AGGR_NODE:
1888                 return perf_stat__get_node_file;
1889         case AGGR_GLOBAL:
1890                 return perf_stat__get_global_file;
1891         case AGGR_NONE:
1892                 return perf_stat__get_cpu_file;
1893         case AGGR_THREAD:
1894         case AGGR_UNSET:
1895         case AGGR_MAX:
1896         default:
1897                 return NULL;
1898         }
1899 }
1900
1901 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1902 {
1903         struct perf_env *env = &st->session->header.env;
1904         aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1905         bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1906
1907         if (stat_config.aggr_mode == AGGR_THREAD) {
1908                 int nr = perf_thread_map__nr(evsel_list->core.threads);
1909
1910                 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1911                 if (stat_config.aggr_map == NULL)
1912                         return -ENOMEM;
1913
1914                 for (int s = 0; s < nr; s++) {
1915                         struct aggr_cpu_id id = aggr_cpu_id__empty();
1916
1917                         id.thread_idx = s;
1918                         stat_config.aggr_map->map[s] = id;
1919                 }
1920                 return 0;
1921         }
1922
1923         if (!get_id)
1924                 return 0;
1925
1926         stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1927                                                  get_id, env, needs_sort);
1928         if (!stat_config.aggr_map) {
1929                 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1930                 return -1;
1931         }
1932         stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1933         return 0;
1934 }
1935
1936 /*
1937  * Add default attributes, if there were no attributes specified or
1938  * if -d/--detailed, -d -d or -d -d -d is used:
1939  */
1940 static int add_default_attributes(void)
1941 {
1942         struct perf_event_attr default_attrs0[] = {
1943
1944   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1945   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1946   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1947   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1948
1949   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1950 };
1951         struct perf_event_attr frontend_attrs[] = {
1952   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1953 };
1954         struct perf_event_attr backend_attrs[] = {
1955   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1956 };
1957         struct perf_event_attr default_attrs1[] = {
1958   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1959   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1960   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1961
1962 };
1963
1964 /*
1965  * Detailed stats (-d), covering the L1 and last level data caches:
1966  */
1967         struct perf_event_attr detailed_attrs[] = {
1968
1969   { .type = PERF_TYPE_HW_CACHE,
1970     .config =
1971          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1972         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1973         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1974
1975   { .type = PERF_TYPE_HW_CACHE,
1976     .config =
1977          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1978         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1979         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1980
1981   { .type = PERF_TYPE_HW_CACHE,
1982     .config =
1983          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1984         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1985         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1986
1987   { .type = PERF_TYPE_HW_CACHE,
1988     .config =
1989          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1990         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1991         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1992 };
1993
1994 /*
1995  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1996  */
1997         struct perf_event_attr very_detailed_attrs[] = {
1998
1999   { .type = PERF_TYPE_HW_CACHE,
2000     .config =
2001          PERF_COUNT_HW_CACHE_L1I                <<  0  |
2002         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2003         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2004
2005   { .type = PERF_TYPE_HW_CACHE,
2006     .config =
2007          PERF_COUNT_HW_CACHE_L1I                <<  0  |
2008         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2009         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2010
2011   { .type = PERF_TYPE_HW_CACHE,
2012     .config =
2013          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2014         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2015         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2016
2017   { .type = PERF_TYPE_HW_CACHE,
2018     .config =
2019          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2020         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2021         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2022
2023   { .type = PERF_TYPE_HW_CACHE,
2024     .config =
2025          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2026         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2027         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2028
2029   { .type = PERF_TYPE_HW_CACHE,
2030     .config =
2031          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2032         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2033         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2034
2035 };
2036
2037 /*
2038  * Very, very detailed stats (-d -d -d), adding prefetch events:
2039  */
2040         struct perf_event_attr very_very_detailed_attrs[] = {
2041
2042   { .type = PERF_TYPE_HW_CACHE,
2043     .config =
2044          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2045         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2046         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2047
2048   { .type = PERF_TYPE_HW_CACHE,
2049     .config =
2050          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2051         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2052         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2053 };
2054
2055         struct perf_event_attr default_null_attrs[] = {};
2056         const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2057
2058         /* Set attrs if no event is selected and !null_run: */
2059         if (stat_config.null_run)
2060                 return 0;
2061
2062         if (transaction_run) {
2063                 /* Handle -T as -M transaction. Once platform specific metrics
2064                  * support has been added to the json files, all architectures
2065                  * will use this approach. To determine transaction support
2066                  * on an architecture test for such a metric name.
2067                  */
2068                 if (!metricgroup__has_metric(pmu, "transaction")) {
2069                         pr_err("Missing transaction metrics\n");
2070                         return -1;
2071                 }
2072                 return metricgroup__parse_groups(evsel_list, pmu, "transaction",
2073                                                 stat_config.metric_no_group,
2074                                                 stat_config.metric_no_merge,
2075                                                 stat_config.metric_no_threshold,
2076                                                 stat_config.user_requested_cpu_list,
2077                                                 stat_config.system_wide,
2078                                                 stat_config.hardware_aware_grouping,
2079                                                 &stat_config.metric_events);
2080         }
2081
2082         if (smi_cost) {
2083                 int smi;
2084
2085                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2086                         pr_err("freeze_on_smi is not supported.\n");
2087                         return -1;
2088                 }
2089
2090                 if (!smi) {
2091                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2092                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2093                                 return -1;
2094                         }
2095                         smi_reset = true;
2096                 }
2097
2098                 if (!metricgroup__has_metric(pmu, "smi")) {
2099                         pr_err("Missing smi metrics\n");
2100                         return -1;
2101                 }
2102
2103                 if (!force_metric_only)
2104                         stat_config.metric_only = true;
2105
2106                 return metricgroup__parse_groups(evsel_list, pmu, "smi",
2107                                                 stat_config.metric_no_group,
2108                                                 stat_config.metric_no_merge,
2109                                                 stat_config.metric_no_threshold,
2110                                                 stat_config.user_requested_cpu_list,
2111                                                 stat_config.system_wide,
2112                                                 stat_config.hardware_aware_grouping,
2113                                                 &stat_config.metric_events);
2114         }
2115
2116         if (topdown_run) {
2117                 unsigned int max_level = metricgroups__topdown_max_level();
2118                 char str[] = "TopdownL1";
2119
2120                 if (!force_metric_only)
2121                         stat_config.metric_only = true;
2122
2123                 if (!max_level) {
2124                         pr_err("Topdown requested but the topdown metric groups aren't present.\n"
2125                                 "(See perf list the metric groups have names like TopdownL1)\n");
2126                         return -1;
2127                 }
2128                 if (stat_config.topdown_level > max_level) {
2129                         pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
2130                         return -1;
2131                 } else if (!stat_config.topdown_level)
2132                         stat_config.topdown_level = 1;
2133
2134                 if (!stat_config.interval && !stat_config.metric_only) {
2135                         fprintf(stat_config.output,
2136                                 "Topdown accuracy may decrease when measuring long periods.\n"
2137                                 "Please print the result regularly, e.g. -I1000\n");
2138                 }
2139                 str[8] = stat_config.topdown_level + '0';
2140                 if (metricgroup__parse_groups(evsel_list,
2141                                                 pmu, str,
2142                                                 /*metric_no_group=*/false,
2143                                                 /*metric_no_merge=*/false,
2144                                                 /*metric_no_threshold=*/true,
2145                                                 stat_config.user_requested_cpu_list,
2146                                                 stat_config.system_wide,
2147                                                 stat_config.hardware_aware_grouping,
2148                                                 &stat_config.metric_events) < 0)
2149                         return -1;
2150         }
2151
2152         if (!stat_config.topdown_level)
2153                 stat_config.topdown_level = 1;
2154
2155         if (!evsel_list->core.nr_entries) {
2156                 /* No events so add defaults. */
2157                 if (target__has_cpu(&target))
2158                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2159
2160                 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2161                         return -1;
2162                 if (perf_pmus__have_event("cpu", "stalled-cycles-frontend")) {
2163                         if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
2164                                 return -1;
2165                 }
2166                 if (perf_pmus__have_event("cpu", "stalled-cycles-backend")) {
2167                         if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2168                                 return -1;
2169                 }
2170                 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2171                         return -1;
2172                 /*
2173                  * Add TopdownL1 metrics if they exist. To minimize
2174                  * multiplexing, don't request threshold computation.
2175                  */
2176                 if (metricgroup__has_metric(pmu, "Default")) {
2177                         struct evlist *metric_evlist = evlist__new();
2178                         struct evsel *metric_evsel;
2179
2180                         if (!metric_evlist)
2181                                 return -1;
2182
2183                         if (metricgroup__parse_groups(metric_evlist, pmu, "Default",
2184                                                         /*metric_no_group=*/false,
2185                                                         /*metric_no_merge=*/false,
2186                                                         /*metric_no_threshold=*/true,
2187                                                         stat_config.user_requested_cpu_list,
2188                                                         stat_config.system_wide,
2189                                                         stat_config.hardware_aware_grouping,
2190                                                         &stat_config.metric_events) < 0)
2191                                 return -1;
2192
2193                         evlist__for_each_entry(metric_evlist, metric_evsel) {
2194                                 metric_evsel->skippable = true;
2195                                 metric_evsel->default_metricgroup = true;
2196                         }
2197                         evlist__splice_list_tail(evsel_list, &metric_evlist->core.entries);
2198                         evlist__delete(metric_evlist);
2199                 }
2200
2201                 /* Platform specific attrs */
2202                 if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2203                         return -1;
2204         }
2205
2206         /* Detailed events get appended to the event list: */
2207
2208         if (detailed_run <  1)
2209                 return 0;
2210
2211         /* Append detailed run extra attributes: */
2212         if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2213                 return -1;
2214
2215         if (detailed_run < 2)
2216                 return 0;
2217
2218         /* Append very detailed run extra attributes: */
2219         if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2220                 return -1;
2221
2222         if (detailed_run < 3)
2223                 return 0;
2224
2225         /* Append very, very detailed run extra attributes: */
2226         return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2227 }
2228
2229 static const char * const stat_record_usage[] = {
2230         "perf stat record [<options>]",
2231         NULL,
2232 };
2233
2234 static void init_features(struct perf_session *session)
2235 {
2236         int feat;
2237
2238         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2239                 perf_header__set_feat(&session->header, feat);
2240
2241         perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2242         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2243         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2244         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2245         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2246 }
2247
2248 static int __cmd_record(int argc, const char **argv)
2249 {
2250         struct perf_session *session;
2251         struct perf_data *data = &perf_stat.data;
2252
2253         argc = parse_options(argc, argv, stat_options, stat_record_usage,
2254                              PARSE_OPT_STOP_AT_NON_OPTION);
2255
2256         if (output_name)
2257                 data->path = output_name;
2258
2259         if (stat_config.run_count != 1 || forever) {
2260                 pr_err("Cannot use -r option with perf stat record.\n");
2261                 return -1;
2262         }
2263
2264         session = perf_session__new(data, NULL);
2265         if (IS_ERR(session)) {
2266                 pr_err("Perf session creation failed\n");
2267                 return PTR_ERR(session);
2268         }
2269
2270         init_features(session);
2271
2272         session->evlist   = evsel_list;
2273         perf_stat.session = session;
2274         perf_stat.record  = true;
2275         return argc;
2276 }
2277
2278 static int process_stat_round_event(struct perf_session *session,
2279                                     union perf_event *event)
2280 {
2281         struct perf_record_stat_round *stat_round = &event->stat_round;
2282         struct timespec tsh, *ts = NULL;
2283         const char **argv = session->header.env.cmdline_argv;
2284         int argc = session->header.env.nr_cmdline;
2285
2286         process_counters();
2287
2288         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2289                 update_stats(&walltime_nsecs_stats, stat_round->time);
2290
2291         if (stat_config.interval && stat_round->time) {
2292                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2293                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2294                 ts = &tsh;
2295         }
2296
2297         print_counters(ts, argc, argv);
2298         return 0;
2299 }
2300
2301 static
2302 int process_stat_config_event(struct perf_session *session,
2303                               union perf_event *event)
2304 {
2305         struct perf_tool *tool = session->tool;
2306         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2307
2308         perf_event__read_stat_config(&stat_config, &event->stat_config);
2309
2310         if (perf_cpu_map__is_empty(st->cpus)) {
2311                 if (st->aggr_mode != AGGR_UNSET)
2312                         pr_warning("warning: processing task data, aggregation mode not set\n");
2313         } else if (st->aggr_mode != AGGR_UNSET) {
2314                 stat_config.aggr_mode = st->aggr_mode;
2315         }
2316
2317         if (perf_stat.data.is_pipe)
2318                 perf_stat_init_aggr_mode();
2319         else
2320                 perf_stat_init_aggr_mode_file(st);
2321
2322         if (stat_config.aggr_map) {
2323                 int nr_aggr = stat_config.aggr_map->nr;
2324
2325                 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2326                         pr_err("cannot allocate aggr counts\n");
2327                         return -1;
2328                 }
2329         }
2330         return 0;
2331 }
2332
2333 static int set_maps(struct perf_stat *st)
2334 {
2335         if (!st->cpus || !st->threads)
2336                 return 0;
2337
2338         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2339                 return -EINVAL;
2340
2341         perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2342
2343         if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2344                 return -ENOMEM;
2345
2346         st->maps_allocated = true;
2347         return 0;
2348 }
2349
2350 static
2351 int process_thread_map_event(struct perf_session *session,
2352                              union perf_event *event)
2353 {
2354         struct perf_tool *tool = session->tool;
2355         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2356
2357         if (st->threads) {
2358                 pr_warning("Extra thread map event, ignoring.\n");
2359                 return 0;
2360         }
2361
2362         st->threads = thread_map__new_event(&event->thread_map);
2363         if (!st->threads)
2364                 return -ENOMEM;
2365
2366         return set_maps(st);
2367 }
2368
2369 static
2370 int process_cpu_map_event(struct perf_session *session,
2371                           union perf_event *event)
2372 {
2373         struct perf_tool *tool = session->tool;
2374         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2375         struct perf_cpu_map *cpus;
2376
2377         if (st->cpus) {
2378                 pr_warning("Extra cpu map event, ignoring.\n");
2379                 return 0;
2380         }
2381
2382         cpus = cpu_map__new_data(&event->cpu_map.data);
2383         if (!cpus)
2384                 return -ENOMEM;
2385
2386         st->cpus = cpus;
2387         return set_maps(st);
2388 }
2389
2390 static const char * const stat_report_usage[] = {
2391         "perf stat report [<options>]",
2392         NULL,
2393 };
2394
2395 static struct perf_stat perf_stat = {
2396         .tool = {
2397                 .attr           = perf_event__process_attr,
2398                 .event_update   = perf_event__process_event_update,
2399                 .thread_map     = process_thread_map_event,
2400                 .cpu_map        = process_cpu_map_event,
2401                 .stat_config    = process_stat_config_event,
2402                 .stat           = perf_event__process_stat_event,
2403                 .stat_round     = process_stat_round_event,
2404         },
2405         .aggr_mode      = AGGR_UNSET,
2406         .aggr_level     = 0,
2407 };
2408
2409 static int __cmd_report(int argc, const char **argv)
2410 {
2411         struct perf_session *session;
2412         const struct option options[] = {
2413         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2414         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2415                      "aggregate counts per processor socket", AGGR_SOCKET),
2416         OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2417                      "aggregate counts per processor die", AGGR_DIE),
2418         OPT_SET_UINT(0, "per-cluster", &perf_stat.aggr_mode,
2419                      "aggregate counts perf processor cluster", AGGR_CLUSTER),
2420         OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
2421                             "cache level",
2422                             "aggregate count at this cache level (Default: LLC)",
2423                             parse_cache_level),
2424         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2425                      "aggregate counts per physical processor core", AGGR_CORE),
2426         OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2427                      "aggregate counts per numa node", AGGR_NODE),
2428         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2429                      "disable CPU count aggregation", AGGR_NONE),
2430         OPT_END()
2431         };
2432         struct stat st;
2433         int ret;
2434
2435         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2436
2437         if (!input_name || !strlen(input_name)) {
2438                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2439                         input_name = "-";
2440                 else
2441                         input_name = "perf.data";
2442         }
2443
2444         perf_stat.data.path = input_name;
2445         perf_stat.data.mode = PERF_DATA_MODE_READ;
2446
2447         session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2448         if (IS_ERR(session))
2449                 return PTR_ERR(session);
2450
2451         perf_stat.session  = session;
2452         stat_config.output = stderr;
2453         evlist__delete(evsel_list);
2454         evsel_list         = session->evlist;
2455
2456         ret = perf_session__process_events(session);
2457         if (ret)
2458                 return ret;
2459
2460         perf_session__delete(session);
2461         return 0;
2462 }
2463
2464 static void setup_system_wide(int forks)
2465 {
2466         /*
2467          * Make system wide (-a) the default target if
2468          * no target was specified and one of following
2469          * conditions is met:
2470          *
2471          *   - there's no workload specified
2472          *   - there is workload specified but all requested
2473          *     events are system wide events
2474          */
2475         if (!target__none(&target))
2476                 return;
2477
2478         if (!forks)
2479                 target.system_wide = true;
2480         else {
2481                 struct evsel *counter;
2482
2483                 evlist__for_each_entry(evsel_list, counter) {
2484                         if (!counter->core.requires_cpu &&
2485                             !evsel__name_is(counter, "duration_time")) {
2486                                 return;
2487                         }
2488                 }
2489
2490                 if (evsel_list->core.nr_entries)
2491                         target.system_wide = true;
2492         }
2493 }
2494
2495 int cmd_stat(int argc, const char **argv)
2496 {
2497         const char * const stat_usage[] = {
2498                 "perf stat [<options>] [<command>]",
2499                 NULL
2500         };
2501         int status = -EINVAL, run_idx, err;
2502         const char *mode;
2503         FILE *output = stderr;
2504         unsigned int interval, timeout;
2505         const char * const stat_subcommands[] = { "record", "report" };
2506         char errbuf[BUFSIZ];
2507
2508         setlocale(LC_ALL, "");
2509
2510         evsel_list = evlist__new();
2511         if (evsel_list == NULL)
2512                 return -ENOMEM;
2513
2514         parse_events__shrink_config_terms();
2515
2516         /* String-parsing callback-based options would segfault when negated */
2517         set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2518         set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2519         set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2520
2521         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2522                                         (const char **) stat_usage,
2523                                         PARSE_OPT_STOP_AT_NON_OPTION);
2524
2525         if (stat_config.csv_sep) {
2526                 stat_config.csv_output = true;
2527                 if (!strcmp(stat_config.csv_sep, "\\t"))
2528                         stat_config.csv_sep = "\t";
2529         } else
2530                 stat_config.csv_sep = DEFAULT_SEPARATOR;
2531
2532         if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2533                 argc = __cmd_record(argc, argv);
2534                 if (argc < 0)
2535                         return -1;
2536         } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2537                 return __cmd_report(argc, argv);
2538
2539         interval = stat_config.interval;
2540         timeout = stat_config.timeout;
2541
2542         /*
2543          * For record command the -o is already taken care of.
2544          */
2545         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2546                 output = NULL;
2547
2548         if (output_name && output_fd) {
2549                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2550                 parse_options_usage(stat_usage, stat_options, "o", 1);
2551                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2552                 goto out;
2553         }
2554
2555         if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2556                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2557                 goto out;
2558         }
2559
2560         if (stat_config.metric_only && stat_config.run_count > 1) {
2561                 fprintf(stderr, "--metric-only is not supported with -r\n");
2562                 goto out;
2563         }
2564
2565         if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2566                 fprintf(stderr, "--table is only supported with -r\n");
2567                 parse_options_usage(stat_usage, stat_options, "r", 1);
2568                 parse_options_usage(NULL, stat_options, "table", 0);
2569                 goto out;
2570         }
2571
2572         if (output_fd < 0) {
2573                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2574                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2575                 goto out;
2576         }
2577
2578         if (!output && !quiet) {
2579                 struct timespec tm;
2580                 mode = append_file ? "a" : "w";
2581
2582                 output = fopen(output_name, mode);
2583                 if (!output) {
2584                         perror("failed to create output file");
2585                         return -1;
2586                 }
2587                 if (!stat_config.json_output) {
2588                         clock_gettime(CLOCK_REALTIME, &tm);
2589                         fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2590                 }
2591         } else if (output_fd > 0) {
2592                 mode = append_file ? "a" : "w";
2593                 output = fdopen(output_fd, mode);
2594                 if (!output) {
2595                         perror("Failed opening logfd");
2596                         return -errno;
2597                 }
2598         }
2599
2600         if (stat_config.interval_clear && !isatty(fileno(output))) {
2601                 fprintf(stderr, "--interval-clear does not work with output\n");
2602                 parse_options_usage(stat_usage, stat_options, "o", 1);
2603                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2604                 parse_options_usage(NULL, stat_options, "interval-clear", 0);
2605                 return -1;
2606         }
2607
2608         stat_config.output = output;
2609
2610         /*
2611          * let the spreadsheet do the pretty-printing
2612          */
2613         if (stat_config.csv_output) {
2614                 /* User explicitly passed -B? */
2615                 if (big_num_opt == 1) {
2616                         fprintf(stderr, "-B option not supported with -x\n");
2617                         parse_options_usage(stat_usage, stat_options, "B", 1);
2618                         parse_options_usage(NULL, stat_options, "x", 1);
2619                         goto out;
2620                 } else /* Nope, so disable big number formatting */
2621                         stat_config.big_num = false;
2622         } else if (big_num_opt == 0) /* User passed --no-big-num */
2623                 stat_config.big_num = false;
2624
2625         err = target__validate(&target);
2626         if (err) {
2627                 target__strerror(&target, err, errbuf, BUFSIZ);
2628                 pr_warning("%s\n", errbuf);
2629         }
2630
2631         setup_system_wide(argc);
2632
2633         /*
2634          * Display user/system times only for single
2635          * run and when there's specified tracee.
2636          */
2637         if ((stat_config.run_count == 1) && target__none(&target))
2638                 stat_config.ru_display = true;
2639
2640         if (stat_config.run_count < 0) {
2641                 pr_err("Run count must be a positive number\n");
2642                 parse_options_usage(stat_usage, stat_options, "r", 1);
2643                 goto out;
2644         } else if (stat_config.run_count == 0) {
2645                 forever = true;
2646                 stat_config.run_count = 1;
2647         }
2648
2649         if (stat_config.walltime_run_table) {
2650                 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2651                 if (!stat_config.walltime_run) {
2652                         pr_err("failed to setup -r option");
2653                         goto out;
2654                 }
2655         }
2656
2657         if ((stat_config.aggr_mode == AGGR_THREAD) &&
2658                 !target__has_task(&target)) {
2659                 if (!target.system_wide || target.cpu_list) {
2660                         fprintf(stderr, "The --per-thread option is only "
2661                                 "available when monitoring via -p -t -a "
2662                                 "options or only --per-thread.\n");
2663                         parse_options_usage(NULL, stat_options, "p", 1);
2664                         parse_options_usage(NULL, stat_options, "t", 1);
2665                         goto out;
2666                 }
2667         }
2668
2669         /*
2670          * no_aggr, cgroup are for system-wide only
2671          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2672          */
2673         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2674               stat_config.aggr_mode != AGGR_THREAD) ||
2675              (nr_cgroups || stat_config.cgroup_list)) &&
2676             !target__has_cpu(&target)) {
2677                 fprintf(stderr, "both cgroup and no-aggregation "
2678                         "modes only available in system-wide mode\n");
2679
2680                 parse_options_usage(stat_usage, stat_options, "G", 1);
2681                 parse_options_usage(NULL, stat_options, "A", 1);
2682                 parse_options_usage(NULL, stat_options, "a", 1);
2683                 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2684                 goto out;
2685         }
2686
2687         if (stat_config.iostat_run) {
2688                 status = iostat_prepare(evsel_list, &stat_config);
2689                 if (status)
2690                         goto out;
2691                 if (iostat_mode == IOSTAT_LIST) {
2692                         iostat_list(evsel_list, &stat_config);
2693                         goto out;
2694                 } else if (verbose > 0)
2695                         iostat_list(evsel_list, &stat_config);
2696                 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2697                         target.system_wide = true;
2698         }
2699
2700         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2701                 target.per_thread = true;
2702
2703         stat_config.system_wide = target.system_wide;
2704         if (target.cpu_list) {
2705                 stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2706                 if (!stat_config.user_requested_cpu_list) {
2707                         status = -ENOMEM;
2708                         goto out;
2709                 }
2710         }
2711
2712         /*
2713          * Metric parsing needs to be delayed as metrics may optimize events
2714          * knowing the target is system-wide.
2715          */
2716         if (metrics) {
2717                 const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2718                 int ret = metricgroup__parse_groups(evsel_list, pmu, metrics,
2719                                                 stat_config.metric_no_group,
2720                                                 stat_config.metric_no_merge,
2721                                                 stat_config.metric_no_threshold,
2722                                                 stat_config.user_requested_cpu_list,
2723                                                 stat_config.system_wide,
2724                                                 stat_config.hardware_aware_grouping,
2725                                                 &stat_config.metric_events);
2726
2727                 zfree(&metrics);
2728                 if (ret) {
2729                         status = ret;
2730                         goto out;
2731                 }
2732         }
2733
2734         if (add_default_attributes())
2735                 goto out;
2736
2737         if (stat_config.cgroup_list) {
2738                 if (nr_cgroups > 0) {
2739                         pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2740                         parse_options_usage(stat_usage, stat_options, "G", 1);
2741                         parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2742                         goto out;
2743                 }
2744
2745                 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2746                                           &stat_config.metric_events, true) < 0) {
2747                         parse_options_usage(stat_usage, stat_options,
2748                                             "for-each-cgroup", 0);
2749                         goto out;
2750                 }
2751         }
2752
2753         evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
2754
2755         if (evlist__create_maps(evsel_list, &target) < 0) {
2756                 if (target__has_task(&target)) {
2757                         pr_err("Problems finding threads of monitor\n");
2758                         parse_options_usage(stat_usage, stat_options, "p", 1);
2759                         parse_options_usage(NULL, stat_options, "t", 1);
2760                 } else if (target__has_cpu(&target)) {
2761                         perror("failed to parse CPUs map");
2762                         parse_options_usage(stat_usage, stat_options, "C", 1);
2763                         parse_options_usage(NULL, stat_options, "a", 1);
2764                 }
2765                 goto out;
2766         }
2767
2768         evlist__check_cpu_maps(evsel_list);
2769
2770         /*
2771          * Initialize thread_map with comm names,
2772          * so we could print it out on output.
2773          */
2774         if (stat_config.aggr_mode == AGGR_THREAD) {
2775                 thread_map__read_comms(evsel_list->core.threads);
2776         }
2777
2778         if (stat_config.aggr_mode == AGGR_NODE)
2779                 cpu__setup_cpunode_map();
2780
2781         if (stat_config.times && interval)
2782                 interval_count = true;
2783         else if (stat_config.times && !interval) {
2784                 pr_err("interval-count option should be used together with "
2785                                 "interval-print.\n");
2786                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2787                 parse_options_usage(stat_usage, stat_options, "I", 1);
2788                 goto out;
2789         }
2790
2791         if (timeout && timeout < 100) {
2792                 if (timeout < 10) {
2793                         pr_err("timeout must be >= 10ms.\n");
2794                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
2795                         goto out;
2796                 } else
2797                         pr_warning("timeout < 100ms. "
2798                                    "The overhead percentage could be high in some cases. "
2799                                    "Please proceed with caution.\n");
2800         }
2801         if (timeout && interval) {
2802                 pr_err("timeout option is not supported with interval-print.\n");
2803                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2804                 parse_options_usage(stat_usage, stat_options, "I", 1);
2805                 goto out;
2806         }
2807
2808         if (perf_stat_init_aggr_mode())
2809                 goto out;
2810
2811         if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2812                 goto out;
2813
2814         /*
2815          * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2816          * while avoiding that older tools show confusing messages.
2817          *
2818          * However for pipe sessions we need to keep it zero,
2819          * because script's perf_evsel__check_attr is triggered
2820          * by attr->sample_type != 0, and we can't run it on
2821          * stat sessions.
2822          */
2823         stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2824
2825         /*
2826          * We dont want to block the signals - that would cause
2827          * child tasks to inherit that and Ctrl-C would not work.
2828          * What we want is for Ctrl-C to work in the exec()-ed
2829          * task, but being ignored by perf stat itself:
2830          */
2831         atexit(sig_atexit);
2832         if (!forever)
2833                 signal(SIGINT,  skip_signal);
2834         signal(SIGCHLD, skip_signal);
2835         signal(SIGALRM, skip_signal);
2836         signal(SIGABRT, skip_signal);
2837
2838         if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2839                 goto out;
2840
2841         /* Enable ignoring missing threads when -p option is defined. */
2842         evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2843         status = 0;
2844         for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2845                 if (stat_config.run_count != 1 && verbose > 0)
2846                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2847                                 run_idx + 1);
2848
2849                 if (run_idx != 0)
2850                         evlist__reset_prev_raw_counts(evsel_list);
2851
2852                 status = run_perf_stat(argc, argv, run_idx);
2853                 if (forever && status != -1 && !interval) {
2854                         print_counters(NULL, argc, argv);
2855                         perf_stat__reset_stats();
2856                 }
2857         }
2858
2859         if (!forever && status != -1 && (!interval || stat_config.summary)) {
2860                 if (stat_config.run_count > 1)
2861                         evlist__copy_res_stats(&stat_config, evsel_list);
2862                 print_counters(NULL, argc, argv);
2863         }
2864
2865         evlist__finalize_ctlfd(evsel_list);
2866
2867         if (STAT_RECORD) {
2868                 /*
2869                  * We synthesize the kernel mmap record just so that older tools
2870                  * don't emit warnings about not being able to resolve symbols
2871                  * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2872                  * a saner message about no samples being in the perf.data file.
2873                  *
2874                  * This also serves to suppress a warning about f_header.data.size == 0
2875                  * in header.c at the moment 'perf stat record' gets introduced, which
2876                  * is not really needed once we start adding the stat specific PERF_RECORD_
2877                  * records, but the need to suppress the kptr_restrict messages in older
2878                  * tools remain  -acme
2879                  */
2880                 int fd = perf_data__fd(&perf_stat.data);
2881
2882                 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2883                                                          process_synthesized_event,
2884                                                          &perf_stat.session->machines.host);
2885                 if (err) {
2886                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2887                                    "older tools may produce warnings about this file\n.");
2888                 }
2889
2890                 if (!interval) {
2891                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2892                                 pr_err("failed to write stat round event\n");
2893                 }
2894
2895                 if (!perf_stat.data.is_pipe) {
2896                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2897                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2898                 }
2899
2900                 evlist__close(evsel_list);
2901                 perf_session__delete(perf_stat.session);
2902         }
2903
2904         perf_stat__exit_aggr_mode();
2905         evlist__free_stats(evsel_list);
2906 out:
2907         if (stat_config.iostat_run)
2908                 iostat_release(evsel_list);
2909
2910         zfree(&stat_config.walltime_run);
2911         zfree(&stat_config.user_requested_cpu_list);
2912
2913         if (smi_cost && smi_reset)
2914                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2915
2916         evlist__delete(evsel_list);
2917
2918         metricgroup__rblist_exit(&stat_config.metric_events);
2919         evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2920
2921         return status;
2922 }