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 <mingo@redhat.com>
  32  *
  33  * Improvements and fixes by:
  34  *
  35  *   Arjan van de Ven <arjan@linux.intel.com>
  36  *   Yanmin Zhang <yanmin.zhang@intel.com>
  37  *   Wu Fengguang <fengguang.wu@intel.com>
  38  *   Mike Galbraith <efault@gmx.de>
  39  *   Paul Mackerras <paulus@samba.org>
  40  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  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 bool cpus_map_matched(struct evsel *a, struct evsel *b)
 168 {
 169         if (!a->core.cpus && !b->core.cpus)
 170                 return true;
 171
 172         if (!a->core.cpus || !b->core.cpus)
 173                 return false;
 174
 175         if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
 176                 return false;
 177
 178         for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
 179                 if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
 180                     perf_cpu_map__cpu(b->core.cpus, i).cpu)
 181                         return false;
 182         }
 183
 184         return true;
 185 }
 186
 187 static void evlist__check_cpu_maps(struct evlist *evlist)
 188 {
 189         struct evsel *evsel, *warned_leader = NULL;
 190
 191         evlist__for_each_entry(evlist, evsel) {
 192                 struct evsel *leader = evsel__leader(evsel);
 193
 194                 /* Check that leader matches cpus with each member. */
 195                 if (leader == evsel)
 196                         continue;
 197                 if (cpus_map_matched(leader, evsel))
 198                         continue;
 199
 200                 /* If there's mismatch disable the group and warn user. */
 201                 if (warned_leader != leader) {
 202                         char buf[200];
 203
 204                         pr_warning("WARNING: grouped events cpus do not match.\n"
 205                                 "Events with CPUs not matching the leader will "
 206                                 "be removed from the group.\n");
 207                         evsel__group_desc(leader, buf, sizeof(buf));
 208                         pr_warning("  %s\n", buf);
 209                         warned_leader = leader;
 210                 }
 211                 if (verbose > 0) {
 212                         char buf[200];
 213
 214                         cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
 215                         pr_warning("     %s: %s\n", leader->name, buf);
 216                         cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
 217                         pr_warning("     %s: %s\n", evsel->name, buf);
 218                 }
 219
 220                 evsel__remove_from_group(evsel, leader);
 221         }
 222 }
 223
 224 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 225                                  struct timespec *b)
 226 {
 227         r->tv_sec = a->tv_sec - b->tv_sec;
 228         if (a->tv_nsec < b->tv_nsec) {
 229                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 230                 r->tv_sec--;
 231         } else {
 232                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 233         }
 234 }
 235
 236 static void perf_stat__reset_stats(void)
 237 {
 238         evlist__reset_stats(evsel_list);
 239         perf_stat__reset_shadow_stats();
 240 }
 241
 242 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 243                                      union perf_event *event,
 244                                      struct perf_sample *sample __maybe_unused,
 245                                      struct machine *machine __maybe_unused)
 246 {
 247         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 248                 pr_err("failed to write perf data, error: %m\n");
 249                 return -1;
 250         }
 251
 252         perf_stat.bytes_written += event->header.size;
 253         return 0;
 254 }
 255
 256 static int write_stat_round_event(u64 tm, u64 type)
 257 {
 258         return perf_event__synthesize_stat_round(NULL, tm, type,
 259                                                  process_synthesized_event,
 260                                                  NULL);
 261 }
 262
 263 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 264         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 265
 266 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
 267
 268 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
 269                                    struct perf_counts_values *count)
 270 {
 271         struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
 272         struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
 273
 274         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 275                                            process_synthesized_event, NULL);
 276 }
 277
 278 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
 279                                int thread, struct timespec *rs)
 280 {
 281         switch(counter->tool_event) {
 282                 case PERF_TOOL_DURATION_TIME: {
 283                         u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
 284                         struct perf_counts_values *count =
 285                                 perf_counts(counter->counts, cpu_map_idx, thread);
 286                         count->ena = count->run = val;
 287                         count->val = val;
 288                         return 0;
 289                 }
 290                 case PERF_TOOL_USER_TIME:
 291                 case PERF_TOOL_SYSTEM_TIME: {
 292                         u64 val;
 293                         struct perf_counts_values *count =
 294                                 perf_counts(counter->counts, cpu_map_idx, thread);
 295                         if (counter->tool_event == PERF_TOOL_USER_TIME)
 296                                 val = ru_stats.ru_utime_usec_stat.mean;
 297                         else
 298                                 val = ru_stats.ru_stime_usec_stat.mean;
 299                         count->ena = count->run = val;
 300                         count->val = val;
 301                         return 0;
 302                 }
 303                 default:
 304                 case PERF_TOOL_NONE:
 305                         return evsel__read_counter(counter, cpu_map_idx, thread);
 306                 case PERF_TOOL_MAX:
 307                         /* This should never be reached */
 308                         return 0;
 309         }
 310 }
 311
 312 /*
 313  * Read out the results of a single counter:
 314  * do not aggregate counts across CPUs in system-wide mode
 315  */
 316 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
 317 {
 318         int nthreads = perf_thread_map__nr(evsel_list->core.threads);
 319         int thread;
 320
 321         if (!counter->supported)
 322                 return -ENOENT;
 323
 324         for (thread = 0; thread < nthreads; thread++) {
 325                 struct perf_counts_values *count;
 326
 327                 count = perf_counts(counter->counts, cpu_map_idx, thread);
 328
 329                 /*
 330                  * The leader's group read loads data into its group members
 331                  * (via evsel__read_counter()) and sets their count->loaded.
 332                  */
 333                 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
 334                     read_single_counter(counter, cpu_map_idx, thread, rs)) {
 335                         counter->counts->scaled = -1;
 336                         perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
 337                         perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
 338                         return -1;
 339                 }
 340
 341                 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
 342
 343                 if (STAT_RECORD) {
 344                         if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
 345                                 pr_err("failed to write stat event\n");
 346                                 return -1;
 347                         }
 348                 }
 349
 350                 if (verbose > 1) {
 351                         fprintf(stat_config.output,
 352                                 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 353                                         evsel__name(counter),
 354                                         perf_cpu_map__cpu(evsel__cpus(counter),
 355                                                           cpu_map_idx).cpu,
 356                                         count->val, count->ena, count->run);
 357                 }
 358         }
 359
 360         return 0;
 361 }
 362
 363 static int read_affinity_counters(struct timespec *rs)
 364 {
 365         struct evlist_cpu_iterator evlist_cpu_itr;
 366         struct affinity saved_affinity, *affinity;
 367
 368         if (all_counters_use_bpf)
 369                 return 0;
 370
 371         if (!target__has_cpu(&target) || target__has_per_thread(&target))
 372                 affinity = NULL;
 373         else if (affinity__setup(&saved_affinity) < 0)
 374                 return -1;
 375         else
 376                 affinity = &saved_affinity;
 377
 378         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 379                 struct evsel *counter = evlist_cpu_itr.evsel;
 380
 381                 if (evsel__is_bpf(counter))
 382                         continue;
 383
 384                 if (!counter->err) {
 385                         counter->err = read_counter_cpu(counter, rs,
 386                                                         evlist_cpu_itr.cpu_map_idx);
 387                 }
 388         }
 389         if (affinity)
 390                 affinity__cleanup(&saved_affinity);
 391
 392         return 0;
 393 }
 394
 395 static int read_bpf_map_counters(void)
 396 {
 397         struct evsel *counter;
 398         int err;
 399
 400         evlist__for_each_entry(evsel_list, counter) {
 401                 if (!evsel__is_bpf(counter))
 402                         continue;
 403
 404                 err = bpf_counter__read(counter);
 405                 if (err)
 406                         return err;
 407         }
 408         return 0;
 409 }
 410
 411 static int read_counters(struct timespec *rs)
 412 {
 413         if (!stat_config.stop_read_counter) {
 414                 if (read_bpf_map_counters() ||
 415                     read_affinity_counters(rs))
 416                         return -1;
 417         }
 418         return 0;
 419 }
 420
 421 static void process_counters(void)
 422 {
 423         struct evsel *counter;
 424
 425         evlist__for_each_entry(evsel_list, counter) {
 426                 if (counter->err)
 427                         pr_debug("failed to read counter %s\n", counter->name);
 428                 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
 429                         pr_warning("failed to process counter %s\n", counter->name);
 430                 counter->err = 0;
 431         }
 432
 433         perf_stat_merge_counters(&stat_config, evsel_list);
 434         perf_stat_process_percore(&stat_config, evsel_list);
 435 }
 436
 437 static void process_interval(void)
 438 {
 439         struct timespec ts, rs;
 440
 441         clock_gettime(CLOCK_MONOTONIC, &ts);
 442         diff_timespec(&rs, &ts, &ref_time);
 443
 444         evlist__reset_aggr_stats(evsel_list);
 445
 446         if (read_counters(&rs) == 0)
 447                 process_counters();
 448
 449         if (STAT_RECORD) {
 450                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 451                         pr_err("failed to write stat round event\n");
 452         }
 453
 454         init_stats(&walltime_nsecs_stats);
 455         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
 456         print_counters(&rs, 0, NULL);
 457 }
 458
 459 static bool handle_interval(unsigned int interval, int *times)
 460 {
 461         if (interval) {
 462                 process_interval();
 463                 if (interval_count && !(--(*times)))
 464                         return true;
 465         }
 466         return false;
 467 }
 468
 469 static int enable_counters(void)
 470 {
 471         struct evsel *evsel;
 472         int err;
 473
 474         evlist__for_each_entry(evsel_list, evsel) {
 475                 if (!evsel__is_bpf(evsel))
 476                         continue;
 477
 478                 err = bpf_counter__enable(evsel);
 479                 if (err)
 480                         return err;
 481         }
 482
 483         if (!target__enable_on_exec(&target)) {
 484                 if (!all_counters_use_bpf)
 485                         evlist__enable(evsel_list);
 486         }
 487         return 0;
 488 }
 489
 490 static void disable_counters(void)
 491 {
 492         struct evsel *counter;
 493
 494         /*
 495          * If we don't have tracee (attaching to task or cpu), counters may
 496          * still be running. To get accurate group ratios, we must stop groups
 497          * from counting before reading their constituent counters.
 498          */
 499         if (!target__none(&target)) {
 500                 evlist__for_each_entry(evsel_list, counter)
 501                         bpf_counter__disable(counter);
 502                 if (!all_counters_use_bpf)
 503                         evlist__disable(evsel_list);
 504         }
 505 }
 506
 507 static volatile sig_atomic_t workload_exec_errno;
 508
 509 /*
 510  * evlist__prepare_workload will send a SIGUSR1
 511  * if the fork fails, since we asked by setting its
 512  * want_signal to true.
 513  */
 514 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 515                                         void *ucontext __maybe_unused)
 516 {
 517         workload_exec_errno = info->si_value.sival_int;
 518 }
 519
 520 static bool evsel__should_store_id(struct evsel *counter)
 521 {
 522         return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
 523 }
 524
 525 static bool is_target_alive(struct target *_target,
 526                             struct perf_thread_map *threads)
 527 {
 528         struct stat st;
 529         int i;
 530
 531         if (!target__has_task(_target))
 532                 return true;
 533
 534         for (i = 0; i < threads->nr; i++) {
 535                 char path[PATH_MAX];
 536
 537                 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
 538                           threads->map[i].pid);
 539
 540                 if (!stat(path, &st))
 541                         return true;
 542         }
 543
 544         return false;
 545 }
 546
 547 static void process_evlist(struct evlist *evlist, unsigned int interval)
 548 {
 549         enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
 550
 551         if (evlist__ctlfd_process(evlist, &cmd) > 0) {
 552                 switch (cmd) {
 553                 case EVLIST_CTL_CMD_ENABLE:
 554                         fallthrough;
 555                 case EVLIST_CTL_CMD_DISABLE:
 556                         if (interval)
 557                                 process_interval();
 558                         break;
 559                 case EVLIST_CTL_CMD_SNAPSHOT:
 560                 case EVLIST_CTL_CMD_ACK:
 561                 case EVLIST_CTL_CMD_UNSUPPORTED:
 562                 case EVLIST_CTL_CMD_EVLIST:
 563                 case EVLIST_CTL_CMD_STOP:
 564                 case EVLIST_CTL_CMD_PING:
 565                 default:
 566                         break;
 567                 }
 568         }
 569 }
 570
 571 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
 572                         int *time_to_sleep)
 573 {
 574         int tts = *time_to_sleep;
 575         struct timespec time_diff;
 576
 577         diff_timespec(&time_diff, time_stop, time_start);
 578
 579         tts -= time_diff.tv_sec * MSEC_PER_SEC +
 580                time_diff.tv_nsec / NSEC_PER_MSEC;
 581
 582         if (tts < 0)
 583                 tts = 0;
 584
 585         *time_to_sleep = tts;
 586 }
 587
 588 static int dispatch_events(bool forks, int timeout, int interval, int *times)
 589 {
 590         int child_exited = 0, status = 0;
 591         int time_to_sleep, sleep_time;
 592         struct timespec time_start, time_stop;
 593
 594         if (interval)
 595                 sleep_time = interval;
 596         else if (timeout)
 597                 sleep_time = timeout;
 598         else
 599                 sleep_time = 1000;
 600
 601         time_to_sleep = sleep_time;
 602
 603         while (!done) {
 604                 if (forks)
 605                         child_exited = waitpid(child_pid, &status, WNOHANG);
 606                 else
 607                         child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
 608
 609                 if (child_exited)
 610                         break;
 611
 612                 clock_gettime(CLOCK_MONOTONIC, &time_start);
 613                 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
 614                         if (timeout || handle_interval(interval, times))
 615                                 break;
 616                         time_to_sleep = sleep_time;
 617                 } else { /* fd revent */
 618                         process_evlist(evsel_list, interval);
 619                         clock_gettime(CLOCK_MONOTONIC, &time_stop);
 620                         compute_tts(&time_start, &time_stop, &time_to_sleep);
 621                 }
 622         }
 623
 624         return status;
 625 }
 626
 627 enum counter_recovery {
 628         COUNTER_SKIP,
 629         COUNTER_RETRY,
 630         COUNTER_FATAL,
 631 };
 632
 633 static enum counter_recovery stat_handle_error(struct evsel *counter)
 634 {
 635         char msg[BUFSIZ];
 636         /*
 637          * PPC returns ENXIO for HW counters until 2.6.37
 638          * (behavior changed with commit b0a873e).
 639          */
 640         if (errno == EINVAL || errno == ENOSYS ||
 641             errno == ENOENT || errno == EOPNOTSUPP ||
 642             errno == ENXIO) {
 643                 if (verbose > 0)
 644                         ui__warning("%s event is not supported by the kernel.\n",
 645                                     evsel__name(counter));
 646                 counter->supported = false;
 647                 /*
 648                  * errored is a sticky flag that means one of the counter's
 649                  * cpu event had a problem and needs to be reexamined.
 650                  */
 651                 counter->errored = true;
 652
 653                 if ((evsel__leader(counter) != counter) ||
 654                     !(counter->core.leader->nr_members > 1))
 655                         return COUNTER_SKIP;
 656         } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
 657                 if (verbose > 0)
 658                         ui__warning("%s\n", msg);
 659                 return COUNTER_RETRY;
 660         } else if (target__has_per_thread(&target) &&
 661                    evsel_list->core.threads &&
 662                    evsel_list->core.threads->err_thread != -1) {
 663                 /*
 664                  * For global --per-thread case, skip current
 665                  * error thread.
 666                  */
 667                 if (!thread_map__remove(evsel_list->core.threads,
 668                                         evsel_list->core.threads->err_thread)) {
 669                         evsel_list->core.threads->err_thread = -1;
 670                         return COUNTER_RETRY;
 671                 }
 672         } else if (counter->skippable) {
 673                 if (verbose > 0)
 674                         ui__warning("skipping event %s that kernel failed to open .\n",
 675                                     evsel__name(counter));
 676                 counter->supported = false;
 677                 counter->errored = true;
 678                 return COUNTER_SKIP;
 679         }
 680
 681         evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
 682         ui__error("%s\n", msg);
 683
 684         if (child_pid != -1)
 685                 kill(child_pid, SIGTERM);
 686         return COUNTER_FATAL;
 687 }
 688
 689 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 690 {
 691         int interval = stat_config.interval;
 692         int times = stat_config.times;
 693         int timeout = stat_config.timeout;
 694         char msg[BUFSIZ];
 695         unsigned long long t0, t1;
 696         struct evsel *counter;
 697         size_t l;
 698         int status = 0;
 699         const bool forks = (argc > 0);
 700         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 701         struct evlist_cpu_iterator evlist_cpu_itr;
 702         struct affinity saved_affinity, *affinity = NULL;
 703         int err;
 704         bool second_pass = false;
 705
 706         if (forks) {
 707                 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
 708                         perror("failed to prepare workload");
 709                         return -1;
 710                 }
 711                 child_pid = evsel_list->workload.pid;
 712         }
 713
 714         if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
 715                 if (affinity__setup(&saved_affinity) < 0)
 716                         return -1;
 717                 affinity = &saved_affinity;
 718         }
 719
 720         evlist__for_each_entry(evsel_list, counter) {
 721                 counter->reset_group = false;
 722                 if (bpf_counter__load(counter, &target))
 723                         return -1;
 724                 if (!(evsel__is_bperf(counter)))
 725                         all_counters_use_bpf = false;
 726         }
 727
 728         evlist__reset_aggr_stats(evsel_list);
 729
 730         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 731                 counter = evlist_cpu_itr.evsel;
 732
 733                 /*
 734                  * bperf calls evsel__open_per_cpu() in bperf__load(), so
 735                  * no need to call it again here.
 736                  */
 737                 if (target.use_bpf)
 738                         break;
 739
 740                 if (counter->reset_group || counter->errored)
 741                         continue;
 742                 if (evsel__is_bperf(counter))
 743                         continue;
 744 try_again:
 745                 if (create_perf_stat_counter(counter, &stat_config, &target,
 746                                              evlist_cpu_itr.cpu_map_idx) < 0) {
 747
 748                         /*
 749                          * Weak group failed. We cannot just undo this here
 750                          * because earlier CPUs might be in group mode, and the kernel
 751                          * doesn't support mixing group and non group reads. Defer
 752                          * it to later.
 753                          * Don't close here because we're in the wrong affinity.
 754                          */
 755                         if ((errno == EINVAL || errno == EBADF) &&
 756                                 evsel__leader(counter) != counter &&
 757                                 counter->weak_group) {
 758                                 evlist__reset_weak_group(evsel_list, counter, false);
 759                                 assert(counter->reset_group);
 760                                 second_pass = true;
 761                                 continue;
 762                         }
 763
 764                         switch (stat_handle_error(counter)) {
 765                         case COUNTER_FATAL:
 766                                 return -1;
 767                         case COUNTER_RETRY:
 768                                 goto try_again;
 769                         case COUNTER_SKIP:
 770                                 continue;
 771                         default:
 772                                 break;
 773                         }
 774
 775                 }
 776                 counter->supported = true;
 777         }
 778
 779         if (second_pass) {
 780                 /*
 781                  * Now redo all the weak group after closing them,
 782                  * and also close errored counters.
 783                  */
 784
 785                 /* First close errored or weak retry */
 786                 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 787                         counter = evlist_cpu_itr.evsel;
 788
 789                         if (!counter->reset_group && !counter->errored)
 790                                 continue;
 791
 792                         perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
 793                 }
 794                 /* Now reopen weak */
 795                 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 796                         counter = evlist_cpu_itr.evsel;
 797
 798                         if (!counter->reset_group)
 799                                 continue;
 800 try_again_reset:
 801                         pr_debug2("reopening weak %s\n", evsel__name(counter));
 802                         if (create_perf_stat_counter(counter, &stat_config, &target,
 803                                                      evlist_cpu_itr.cpu_map_idx) < 0) {
 804
 805                                 switch (stat_handle_error(counter)) {
 806                                 case COUNTER_FATAL:
 807                                         return -1;
 808                                 case COUNTER_RETRY:
 809                                         goto try_again_reset;
 810                                 case COUNTER_SKIP:
 811                                         continue;
 812                                 default:
 813                                         break;
 814                                 }
 815                         }
 816                         counter->supported = true;
 817                 }
 818         }
 819         affinity__cleanup(affinity);
 820
 821         evlist__for_each_entry(evsel_list, counter) {
 822                 if (!counter->supported) {
 823                         perf_evsel__free_fd(&counter->core);
 824                         continue;
 825                 }
 826
 827                 l = strlen(counter->unit);
 828                 if (l > stat_config.unit_width)
 829                         stat_config.unit_width = l;
 830
 831                 if (evsel__should_store_id(counter) &&
 832                     evsel__store_ids(counter, evsel_list))
 833                         return -1;
 834         }
 835
 836         if (evlist__apply_filters(evsel_list, &counter)) {
 837                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 838                         counter->filter, evsel__name(counter), errno,
 839                         str_error_r(errno, msg, sizeof(msg)));
 840                 return -1;
 841         }
 842
 843         if (STAT_RECORD) {
 844                 int fd = perf_data__fd(&perf_stat.data);
 845
 846                 if (is_pipe) {
 847                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 848                 } else {
 849                         err = perf_session__write_header(perf_stat.session, evsel_list,
 850                                                          fd, false);
 851                 }
 852
 853                 if (err < 0)
 854                         return err;
 855
 856                 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
 857                                                          process_synthesized_event, is_pipe);
 858                 if (err < 0)
 859                         return err;
 860         }
 861
 862         if (target.initial_delay) {
 863                 pr_info(EVLIST_DISABLED_MSG);
 864         } else {
 865                 err = enable_counters();
 866                 if (err)
 867                         return -1;
 868         }
 869
 870         /* Exec the command, if any */
 871         if (forks)
 872                 evlist__start_workload(evsel_list);
 873
 874         if (target.initial_delay > 0) {
 875                 usleep(target.initial_delay * USEC_PER_MSEC);
 876                 err = enable_counters();
 877                 if (err)
 878                         return -1;
 879
 880                 pr_info(EVLIST_ENABLED_MSG);
 881         }
 882
 883         t0 = rdclock();
 884         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 885
 886         if (forks) {
 887                 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
 888                         status = dispatch_events(forks, timeout, interval, &times);
 889                 if (child_pid != -1) {
 890                         if (timeout)
 891                                 kill(child_pid, SIGTERM);
 892                         wait4(child_pid, &status, 0, &stat_config.ru_data);
 893                 }
 894
 895                 if (workload_exec_errno) {
 896                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 897                         pr_err("Workload failed: %s\n", emsg);
 898                         return -1;
 899                 }
 900
 901                 if (WIFSIGNALED(status))
 902                         psignal(WTERMSIG(status), argv[0]);
 903         } else {
 904                 status = dispatch_events(forks, timeout, interval, &times);
 905         }
 906
 907         disable_counters();
 908
 909         t1 = rdclock();
 910
 911         if (stat_config.walltime_run_table)
 912                 stat_config.walltime_run[run_idx] = t1 - t0;
 913
 914         if (interval && stat_config.summary) {
 915                 stat_config.interval = 0;
 916                 stat_config.stop_read_counter = true;
 917                 init_stats(&walltime_nsecs_stats);
 918                 update_stats(&walltime_nsecs_stats, t1 - t0);
 919
 920                 evlist__copy_prev_raw_counts(evsel_list);
 921                 evlist__reset_prev_raw_counts(evsel_list);
 922                 evlist__reset_aggr_stats(evsel_list);
 923         } else {
 924                 update_stats(&walltime_nsecs_stats, t1 - t0);
 925                 update_rusage_stats(&ru_stats, &stat_config.ru_data);
 926         }
 927
 928         /*
 929          * Closing a group leader splits the group, and as we only disable
 930          * group leaders, results in remaining events becoming enabled. To
 931          * avoid arbitrary skew, we must read all counters before closing any
 932          * group leaders.
 933          */
 934         if (read_counters(&(struct timespec) { .tv_nsec = t1-t0 }) == 0)
 935                 process_counters();
 936
 937         /*
 938          * We need to keep evsel_list alive, because it's processed
 939          * later the evsel_list will be closed after.
 940          */
 941         if (!STAT_RECORD)
 942                 evlist__close(evsel_list);
 943
 944         return WEXITSTATUS(status);
 945 }
 946
 947 static int run_perf_stat(int argc, const char **argv, int run_idx)
 948 {
 949         int ret;
 950
 951         if (pre_cmd) {
 952                 ret = system(pre_cmd);
 953                 if (ret)
 954                         return ret;
 955         }
 956
 957         if (sync_run)
 958                 sync();
 959
 960         ret = __run_perf_stat(argc, argv, run_idx);
 961         if (ret)
 962                 return ret;
 963
 964         if (post_cmd) {
 965                 ret = system(post_cmd);
 966                 if (ret)
 967                         return ret;
 968         }
 969
 970         return ret;
 971 }
 972
 973 static void print_counters(struct timespec *ts, int argc, const char **argv)
 974 {
 975         /* Do not print anything if we record to the pipe. */
 976         if (STAT_RECORD && perf_stat.data.is_pipe)
 977                 return;
 978         if (quiet)
 979                 return;
 980
 981         evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
 982 }
 983
 984 static volatile sig_atomic_t signr = -1;
 985
 986 static void skip_signal(int signo)
 987 {
 988         if ((child_pid == -1) || stat_config.interval)
 989                 done = 1;
 990
 991         signr = signo;
 992         /*
 993          * render child_pid harmless
 994          * won't send SIGTERM to a random
 995          * process in case of race condition
 996          * and fast PID recycling
 997          */
 998         child_pid = -1;
 999 }
1000
1001 static void sig_atexit(void)
1002 {
1003         sigset_t set, oset;
1004
1005         /*
1006          * avoid race condition with SIGCHLD handler
1007          * in skip_signal() which is modifying child_pid
1008          * goal is to avoid send SIGTERM to a random
1009          * process
1010          */
1011         sigemptyset(&set);
1012         sigaddset(&set, SIGCHLD);
1013         sigprocmask(SIG_BLOCK, &set, &oset);
1014
1015         if (child_pid != -1)
1016                 kill(child_pid, SIGTERM);
1017
1018         sigprocmask(SIG_SETMASK, &oset, NULL);
1019
1020         if (signr == -1)
1021                 return;
1022
1023         signal(signr, SIG_DFL);
1024         kill(getpid(), signr);
1025 }
1026
1027 void perf_stat__set_big_num(int set)
1028 {
1029         stat_config.big_num = (set != 0);
1030 }
1031
1032 void perf_stat__set_no_csv_summary(int set)
1033 {
1034         stat_config.no_csv_summary = (set != 0);
1035 }
1036
1037 static int stat__set_big_num(const struct option *opt __maybe_unused,
1038                              const char *s __maybe_unused, int unset)
1039 {
1040         big_num_opt = unset ? 0 : 1;
1041         perf_stat__set_big_num(!unset);
1042         return 0;
1043 }
1044
1045 static int enable_metric_only(const struct option *opt __maybe_unused,
1046                               const char *s __maybe_unused, int unset)
1047 {
1048         force_metric_only = true;
1049         stat_config.metric_only = !unset;
1050         return 0;
1051 }
1052
1053 static int append_metric_groups(const struct option *opt __maybe_unused,
1054                                const char *str,
1055                                int unset __maybe_unused)
1056 {
1057         if (metrics) {
1058                 char *tmp;
1059
1060                 if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1061                         return -ENOMEM;
1062                 free(metrics);
1063                 metrics = tmp;
1064         } else {
1065                 metrics = strdup(str);
1066                 if (!metrics)
1067                         return -ENOMEM;
1068         }
1069         return 0;
1070 }
1071
1072 static int parse_control_option(const struct option *opt,
1073                                 const char *str,
1074                                 int unset __maybe_unused)
1075 {
1076         struct perf_stat_config *config = opt->value;
1077
1078         return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1079 }
1080
1081 static int parse_stat_cgroups(const struct option *opt,
1082                               const char *str, int unset)
1083 {
1084         if (stat_config.cgroup_list) {
1085                 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1086                 return -1;
1087         }
1088
1089         return parse_cgroups(opt, str, unset);
1090 }
1091
1092 static int parse_cputype(const struct option *opt,
1093                              const char *str,
1094                              int unset __maybe_unused)
1095 {
1096         const struct perf_pmu *pmu;
1097         struct evlist *evlist = *(struct evlist **)opt->value;
1098
1099         if (!list_empty(&evlist->core.entries)) {
1100                 fprintf(stderr, "Must define cputype before events/metrics\n");
1101                 return -1;
1102         }
1103
1104         pmu = perf_pmus__pmu_for_pmu_filter(str);
1105         if (!pmu) {
1106                 fprintf(stderr, "--cputype %s is not supported!\n", str);
1107                 return -1;
1108         }
1109         parse_events_option_args.pmu_filter = pmu->name;
1110
1111         return 0;
1112 }
1113
1114 static int parse_cache_level(const struct option *opt,
1115                              const char *str,
1116                              int unset __maybe_unused)
1117 {
1118         int level;
1119         u32 *aggr_mode = (u32 *)opt->value;
1120         u32 *aggr_level = (u32 *)opt->data;
1121
1122         /*
1123          * If no string is specified, aggregate based on the topology of
1124          * Last Level Cache (LLC). Since the LLC level can change from
1125          * architecture to architecture, set level greater than
1126          * MAX_CACHE_LVL which will be interpreted as LLC.
1127          */
1128         if (str == NULL) {
1129                 level = MAX_CACHE_LVL + 1;
1130                 goto out;
1131         }
1132
1133         /*
1134          * The format to specify cache level is LX or lX where X is the
1135          * cache level.
1136          */
1137         if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
1138                 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1139                        MAX_CACHE_LVL,
1140                        MAX_CACHE_LVL);
1141                 return -EINVAL;
1142         }
1143
1144         level = atoi(&str[1]);
1145         if (level < 1) {
1146                 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1147                        MAX_CACHE_LVL,
1148                        MAX_CACHE_LVL);
1149                 return -EINVAL;
1150         }
1151
1152         if (level > MAX_CACHE_LVL) {
1153                 pr_err("perf only supports max cache level of %d.\n"
1154                        "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
1155                 return -EINVAL;
1156         }
1157 out:
1158         *aggr_mode = AGGR_CACHE;
1159         *aggr_level = level;
1160         return 0;
1161 }
1162
1163 static struct option stat_options[] = {
1164         OPT_BOOLEAN('T', "transaction", &transaction_run,
1165                     "hardware transaction statistics"),
1166         OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
1167                      "event selector. use 'perf list' to list available events",
1168                      parse_events_option),
1169         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1170                      "event filter", parse_filter),
1171         OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1172                     "child tasks do not inherit counters"),
1173         OPT_STRING('p', "pid", &target.pid, "pid",
1174                    "stat events on existing process id"),
1175         OPT_STRING('t', "tid", &target.tid, "tid",
1176                    "stat events on existing thread id"),
1177 #ifdef HAVE_BPF_SKEL
1178         OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1179                    "stat events on existing bpf program id"),
1180         OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1181                     "use bpf program to count events"),
1182         OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1183                    "path to perf_event_attr map"),
1184 #endif
1185         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1186                     "system-wide collection from all CPUs"),
1187         OPT_BOOLEAN(0, "scale", &stat_config.scale,
1188                     "Use --no-scale to disable counter scaling for multiplexing"),
1189         OPT_INCR('v', "verbose", &verbose,
1190                     "be more verbose (show counter open errors, etc)"),
1191         OPT_INTEGER('r', "repeat", &stat_config.run_count,
1192                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1193         OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1194                     "display details about each run (only with -r option)"),
1195         OPT_BOOLEAN('n', "null", &stat_config.null_run,
1196                     "null run - dont start any counters"),
1197         OPT_INCR('d', "detailed", &detailed_run,
1198                     "detailed run - start a lot of events"),
1199         OPT_BOOLEAN('S', "sync", &sync_run,
1200                     "call sync() before starting a run"),
1201         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1202                            "print large numbers with thousands\' separators",
1203                            stat__set_big_num),
1204         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1205                     "list of cpus to monitor in system-wide"),
1206         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1207                     "disable aggregation across CPUs or PMUs", AGGR_NONE),
1208         OPT_SET_UINT(0, "no-merge", &stat_config.aggr_mode,
1209                     "disable aggregation the same as -A or -no-aggr", AGGR_NONE),
1210         OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1211                     "Merge identical named hybrid events"),
1212         OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1213                    "print counts with custom separator"),
1214         OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1215                    "print counts in JSON format"),
1216         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1217                      "monitor event in cgroup name only", parse_stat_cgroups),
1218         OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1219                     "expand events for each cgroup"),
1220         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1221         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1222         OPT_INTEGER(0, "log-fd", &output_fd,
1223                     "log output to fd, instead of stderr"),
1224         OPT_STRING(0, "pre", &pre_cmd, "command",
1225                         "command to run prior to the measured command"),
1226         OPT_STRING(0, "post", &post_cmd, "command",
1227                         "command to run after to the measured command"),
1228         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1229                     "print counts at regular interval in ms "
1230                     "(overhead is possible for values <= 100ms)"),
1231         OPT_INTEGER(0, "interval-count", &stat_config.times,
1232                     "print counts for fixed number of times"),
1233         OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1234                     "clear screen in between new interval"),
1235         OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1236                     "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1237         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1238                      "aggregate counts per processor socket", AGGR_SOCKET),
1239         OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1240                      "aggregate counts per processor die", AGGR_DIE),
1241         OPT_CALLBACK_OPTARG(0, "per-cache", &stat_config.aggr_mode, &stat_config.aggr_level,
1242                             "cache level", "aggregate count at this cache level (Default: LLC)",
1243                             parse_cache_level),
1244         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1245                      "aggregate counts per physical processor core", AGGR_CORE),
1246         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1247                      "aggregate counts per thread", AGGR_THREAD),
1248         OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1249                      "aggregate counts per numa node", AGGR_NODE),
1250         OPT_INTEGER('D', "delay", &target.initial_delay,
1251                     "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1252         OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1253                         "Only print computed metrics. No raw values", enable_metric_only),
1254         OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1255                        "don't group metric events, impacts multiplexing"),
1256         OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1257                        "don't try to share events between metrics in a group"),
1258         OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold,
1259                        "disable adding events for the metric threshold calculation"),
1260         OPT_BOOLEAN(0, "topdown", &topdown_run,
1261                         "measure top-down statistics"),
1262         OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1263                         "Set the metrics level for the top-down statistics (0: max level)"),
1264         OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1265                         "measure SMI cost"),
1266         OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1267                      "monitor specified metrics or metric groups (separated by ,)",
1268                      append_metric_groups),
1269         OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1270                          "Configure all used events to run in kernel space.",
1271                          PARSE_OPT_EXCLUSIVE),
1272         OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1273                          "Configure all used events to run in user space.",
1274                          PARSE_OPT_EXCLUSIVE),
1275         OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1276                     "Use with 'percore' event qualifier to show the event "
1277                     "counts of one hardware thread by sum up total hardware "
1278                     "threads of same physical core"),
1279         OPT_BOOLEAN(0, "summary", &stat_config.summary,
1280                        "print summary for interval mode"),
1281         OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1282                        "don't print 'summary' for CSV summary output"),
1283         OPT_BOOLEAN(0, "quiet", &quiet,
1284                         "don't print any output, messages or warnings (useful with record)"),
1285         OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1286                      "Only enable events on applying cpu with this type "
1287                      "for hybrid platform (e.g. core or atom)",
1288                      parse_cputype),
1289 #ifdef HAVE_LIBPFM
1290         OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1291                 "libpfm4 event selector. use 'perf list' to list available events",
1292                 parse_libpfm_events_option),
1293 #endif
1294         OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1295                      "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1296                      "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1297                      "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1298                       parse_control_option),
1299         OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1300                             "measure I/O performance metrics provided by arch/platform",
1301                             iostat_parse),
1302         OPT_END()
1303 };
1304
1305 /**
1306  * Calculate the cache instance ID from the map in
1307  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1308  * Cache instance ID is the first CPU reported in the shared_cpu_list file.
1309  */
1310 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
1311 {
1312         int id;
1313         struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
1314
1315         /*
1316          * If the map contains no CPU, consider the current CPU to
1317          * be the first online CPU in the cache domain else use the
1318          * first online CPU of the cache domain as the ID.
1319          */
1320         if (perf_cpu_map__has_any_cpu_or_is_empty(cpu_map))
1321                 id = cpu.cpu;
1322         else
1323                 id = perf_cpu_map__cpu(cpu_map, 0).cpu;
1324
1325         /* Free the perf_cpu_map used to find the cache ID */
1326         perf_cpu_map__put(cpu_map);
1327
1328         return id;
1329 }
1330
1331 /**
1332  * cpu__get_cache_id - Returns 0 if successful in populating the
1333  * cache level and cache id. Cache level is read from
1334  * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
1335  * is the first CPU reported by
1336  * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1337  */
1338 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
1339 {
1340         int ret = 0;
1341         u32 cache_level = stat_config.aggr_level;
1342         struct cpu_cache_level caches[MAX_CACHE_LVL];
1343         u32 i = 0, caches_cnt = 0;
1344
1345         cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1346         cache->cache = -1;
1347
1348         ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
1349         if (ret) {
1350                 /*
1351                  * If caches_cnt is not 0, cpu_cache_level data
1352                  * was allocated when building the topology.
1353                  * Free the allocated data before returning.
1354                  */
1355                 if (caches_cnt)
1356                         goto free_caches;
1357
1358                 return ret;
1359         }
1360
1361         if (!caches_cnt)
1362                 return -1;
1363
1364         /*
1365          * Save the data for the highest level if no
1366          * level was specified by the user.
1367          */
1368         if (cache_level > MAX_CACHE_LVL) {
1369                 int max_level_index = 0;
1370
1371                 for (i = 1; i < caches_cnt; ++i) {
1372                         if (caches[i].level > caches[max_level_index].level)
1373                                 max_level_index = i;
1374                 }
1375
1376                 cache->cache_lvl = caches[max_level_index].level;
1377                 cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
1378
1379                 /* Reset i to 0 to free entire caches[] */
1380                 i = 0;
1381                 goto free_caches;
1382         }
1383
1384         for (i = 0; i < caches_cnt; ++i) {
1385                 if (caches[i].level == cache_level) {
1386                         cache->cache_lvl = cache_level;
1387                         cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1388                 }
1389
1390                 cpu_cache_level__free(&caches[i]);
1391         }
1392
1393 free_caches:
1394         /*
1395          * Free all the allocated cpu_cache_level data.
1396          */
1397         while (i < caches_cnt)
1398                 cpu_cache_level__free(&caches[i++]);
1399
1400         return ret;
1401 }
1402
1403 /**
1404  * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
1405  * level, die and socket populated with the cache instache ID, cache level,
1406  * die and socket for cpu. The function signature is compatible with
1407  * aggr_cpu_id_get_t.
1408  */
1409 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
1410 {
1411         int ret;
1412         struct aggr_cpu_id id;
1413         struct perf_cache cache;
1414
1415         id = aggr_cpu_id__die(cpu, data);
1416         if (aggr_cpu_id__is_empty(&id))
1417                 return id;
1418
1419         ret = cpu__get_cache_details(cpu, &cache);
1420         if (ret)
1421                 return id;
1422
1423         id.cache_lvl = cache.cache_lvl;
1424         id.cache = cache.cache;
1425         return id;
1426 }
1427
1428 static const char *const aggr_mode__string[] = {
1429         [AGGR_CORE] = "core",
1430         [AGGR_CACHE] = "cache",
1431         [AGGR_DIE] = "die",
1432         [AGGR_GLOBAL] = "global",
1433         [AGGR_NODE] = "node",
1434         [AGGR_NONE] = "none",
1435         [AGGR_SOCKET] = "socket",
1436         [AGGR_THREAD] = "thread",
1437         [AGGR_UNSET] = "unset",
1438 };
1439
1440 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1441                                                 struct perf_cpu cpu)
1442 {
1443         return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1444 }
1445
1446 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1447                                              struct perf_cpu cpu)
1448 {
1449         return aggr_cpu_id__die(cpu, /*data=*/NULL);
1450 }
1451
1452 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
1453                                                   struct perf_cpu cpu)
1454 {
1455         return aggr_cpu_id__cache(cpu, /*data=*/NULL);
1456 }
1457
1458 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1459                                               struct perf_cpu cpu)
1460 {
1461         return aggr_cpu_id__core(cpu, /*data=*/NULL);
1462 }
1463
1464 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1465                                               struct perf_cpu cpu)
1466 {
1467         return aggr_cpu_id__node(cpu, /*data=*/NULL);
1468 }
1469
1470 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1471                                                 struct perf_cpu cpu)
1472 {
1473         return aggr_cpu_id__global(cpu, /*data=*/NULL);
1474 }
1475
1476 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1477                                              struct perf_cpu cpu)
1478 {
1479         return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1480 }
1481
1482 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1483                                               aggr_get_id_t get_id, struct perf_cpu cpu)
1484 {
1485         struct aggr_cpu_id id;
1486
1487         /* per-process mode - should use global aggr mode */
1488         if (cpu.cpu == -1)
1489                 return get_id(config, cpu);
1490
1491         if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1492                 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1493
1494         id = config->cpus_aggr_map->map[cpu.cpu];
1495         return id;
1496 }
1497
1498 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1499                                                        struct perf_cpu cpu)
1500 {
1501         return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1502 }
1503
1504 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1505                                                     struct perf_cpu cpu)
1506 {
1507         return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1508 }
1509
1510 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
1511                                                          struct perf_cpu cpu)
1512 {
1513         return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
1514 }
1515
1516 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1517                                                      struct perf_cpu cpu)
1518 {
1519         return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1520 }
1521
1522 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1523                                                      struct perf_cpu cpu)
1524 {
1525         return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1526 }
1527
1528 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1529                                                        struct perf_cpu cpu)
1530 {
1531         return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1532 }
1533
1534 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1535                                                     struct perf_cpu cpu)
1536 {
1537         return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1538 }
1539
1540 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1541 {
1542         switch (aggr_mode) {
1543         case AGGR_SOCKET:
1544                 return aggr_cpu_id__socket;
1545         case AGGR_DIE:
1546                 return aggr_cpu_id__die;
1547         case AGGR_CACHE:
1548                 return aggr_cpu_id__cache;
1549         case AGGR_CORE:
1550                 return aggr_cpu_id__core;
1551         case AGGR_NODE:
1552                 return aggr_cpu_id__node;
1553         case AGGR_NONE:
1554                 return aggr_cpu_id__cpu;
1555         case AGGR_GLOBAL:
1556                 return aggr_cpu_id__global;
1557         case AGGR_THREAD:
1558         case AGGR_UNSET:
1559         case AGGR_MAX:
1560         default:
1561                 return NULL;
1562         }
1563 }
1564
1565 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1566 {
1567         switch (aggr_mode) {
1568         case AGGR_SOCKET:
1569                 return perf_stat__get_socket_cached;
1570         case AGGR_DIE:
1571                 return perf_stat__get_die_cached;
1572         case AGGR_CACHE:
1573                 return perf_stat__get_cache_id_cached;
1574         case AGGR_CORE:
1575                 return perf_stat__get_core_cached;
1576         case AGGR_NODE:
1577                 return perf_stat__get_node_cached;
1578         case AGGR_NONE:
1579                 return perf_stat__get_cpu_cached;
1580         case AGGR_GLOBAL:
1581                 return perf_stat__get_global_cached;
1582         case AGGR_THREAD:
1583         case AGGR_UNSET:
1584         case AGGR_MAX:
1585         default:
1586                 return NULL;
1587         }
1588 }
1589
1590 static int perf_stat_init_aggr_mode(void)
1591 {
1592         int nr;
1593         aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1594
1595         if (get_id) {
1596                 bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1597                 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1598                                                          get_id, /*data=*/NULL, needs_sort);
1599                 if (!stat_config.aggr_map) {
1600                         pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1601                         return -1;
1602                 }
1603                 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1604         }
1605
1606         if (stat_config.aggr_mode == AGGR_THREAD) {
1607                 nr = perf_thread_map__nr(evsel_list->core.threads);
1608                 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1609                 if (stat_config.aggr_map == NULL)
1610                         return -ENOMEM;
1611
1612                 for (int s = 0; s < nr; s++) {
1613                         struct aggr_cpu_id id = aggr_cpu_id__empty();
1614
1615                         id.thread_idx = s;
1616                         stat_config.aggr_map->map[s] = id;
1617                 }
1618                 return 0;
1619         }
1620
1621         /*
1622          * The evsel_list->cpus is the base we operate on,
1623          * taking the highest cpu number to be the size of
1624          * the aggregation translate cpumap.
1625          */
1626         if (!perf_cpu_map__has_any_cpu_or_is_empty(evsel_list->core.user_requested_cpus))
1627                 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1628         else
1629                 nr = 0;
1630         stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1631         return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1632 }
1633
1634 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1635 {
1636         if (map) {
1637                 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1638                           "cpu_aggr_map refcnt unbalanced\n");
1639                 free(map);
1640         }
1641 }
1642
1643 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1644 {
1645         if (map && refcount_dec_and_test(&map->refcnt))
1646                 cpu_aggr_map__delete(map);
1647 }
1648
1649 static void perf_stat__exit_aggr_mode(void)
1650 {
1651         cpu_aggr_map__put(stat_config.aggr_map);
1652         cpu_aggr_map__put(stat_config.cpus_aggr_map);
1653         stat_config.aggr_map = NULL;
1654         stat_config.cpus_aggr_map = NULL;
1655 }
1656
1657 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1658 {
1659         struct perf_env *env = data;
1660         struct aggr_cpu_id id = aggr_cpu_id__empty();
1661
1662         if (cpu.cpu != -1)
1663                 id.socket = env->cpu[cpu.cpu].socket_id;
1664
1665         return id;
1666 }
1667
1668 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1669 {
1670         struct perf_env *env = data;
1671         struct aggr_cpu_id id = aggr_cpu_id__empty();
1672
1673         if (cpu.cpu != -1) {
1674                 /*
1675                  * die_id is relative to socket, so start
1676                  * with the socket ID and then add die to
1677                  * make a unique ID.
1678                  */
1679                 id.socket = env->cpu[cpu.cpu].socket_id;
1680                 id.die = env->cpu[cpu.cpu].die_id;
1681         }
1682
1683         return id;
1684 }
1685
1686 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
1687                                            u32 cache_level, struct aggr_cpu_id *id)
1688 {
1689         int i;
1690         int caches_cnt = env->caches_cnt;
1691         struct cpu_cache_level *caches = env->caches;
1692
1693         id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1694         id->cache = -1;
1695
1696         if (!caches_cnt)
1697                 return;
1698
1699         for (i = caches_cnt - 1; i > -1; --i) {
1700                 struct perf_cpu_map *cpu_map;
1701                 int map_contains_cpu;
1702
1703                 /*
1704                  * If user has not specified a level, find the fist level with
1705                  * the cpu in the map. Since building the map is expensive, do
1706                  * this only if levels match.
1707                  */
1708                 if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
1709                         continue;
1710
1711                 cpu_map = perf_cpu_map__new(caches[i].map);
1712                 map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
1713                 perf_cpu_map__put(cpu_map);
1714
1715                 if (map_contains_cpu != -1) {
1716                         id->cache_lvl = caches[i].level;
1717                         id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1718                         return;
1719                 }
1720         }
1721 }
1722
1723 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
1724                                                           void *data)
1725 {
1726         struct perf_env *env = data;
1727         struct aggr_cpu_id id = aggr_cpu_id__empty();
1728
1729         if (cpu.cpu != -1) {
1730                 u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
1731
1732                 id.socket = env->cpu[cpu.cpu].socket_id;
1733                 id.die = env->cpu[cpu.cpu].die_id;
1734                 perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
1735         }
1736
1737         return id;
1738 }
1739
1740 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1741 {
1742         struct perf_env *env = data;
1743         struct aggr_cpu_id id = aggr_cpu_id__empty();
1744
1745         if (cpu.cpu != -1) {
1746                 /*
1747                  * core_id is relative to socket and die,
1748                  * we need a global id. So we set
1749                  * socket, die id and core id
1750                  */
1751                 id.socket = env->cpu[cpu.cpu].socket_id;
1752                 id.die = env->cpu[cpu.cpu].die_id;
1753                 id.core = env->cpu[cpu.cpu].core_id;
1754         }
1755
1756         return id;
1757 }
1758
1759 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1760 {
1761         struct perf_env *env = data;
1762         struct aggr_cpu_id id = aggr_cpu_id__empty();
1763
1764         if (cpu.cpu != -1) {
1765                 /*
1766                  * core_id is relative to socket and die,
1767                  * we need a global id. So we set
1768                  * socket, die id and core id
1769                  */
1770                 id.socket = env->cpu[cpu.cpu].socket_id;
1771                 id.die = env->cpu[cpu.cpu].die_id;
1772                 id.core = env->cpu[cpu.cpu].core_id;
1773                 id.cpu = cpu;
1774         }
1775
1776         return id;
1777 }
1778
1779 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1780 {
1781         struct aggr_cpu_id id = aggr_cpu_id__empty();
1782
1783         id.node = perf_env__numa_node(data, cpu);
1784         return id;
1785 }
1786
1787 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1788                                                            void *data __maybe_unused)
1789 {
1790         struct aggr_cpu_id id = aggr_cpu_id__empty();
1791
1792         /* it always aggregates to the cpu 0 */
1793         id.cpu = (struct perf_cpu){ .cpu = 0 };
1794         return id;
1795 }
1796
1797 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1798                                                      struct perf_cpu cpu)
1799 {
1800         return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1801 }
1802 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1803                                                   struct perf_cpu cpu)
1804 {
1805         return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1806 }
1807
1808 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
1809                                                     struct perf_cpu cpu)
1810 {
1811         return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1812 }
1813
1814 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1815                                                    struct perf_cpu cpu)
1816 {
1817         return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1818 }
1819
1820 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1821                                                   struct perf_cpu cpu)
1822 {
1823         return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1824 }
1825
1826 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1827                                                    struct perf_cpu cpu)
1828 {
1829         return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1830 }
1831
1832 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1833                                                      struct perf_cpu cpu)
1834 {
1835         return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1836 }
1837
1838 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1839 {
1840         switch (aggr_mode) {
1841         case AGGR_SOCKET:
1842                 return perf_env__get_socket_aggr_by_cpu;
1843         case AGGR_DIE:
1844                 return perf_env__get_die_aggr_by_cpu;
1845         case AGGR_CACHE:
1846                 return perf_env__get_cache_aggr_by_cpu;
1847         case AGGR_CORE:
1848                 return perf_env__get_core_aggr_by_cpu;
1849         case AGGR_NODE:
1850                 return perf_env__get_node_aggr_by_cpu;
1851         case AGGR_GLOBAL:
1852                 return perf_env__get_global_aggr_by_cpu;
1853         case AGGR_NONE:
1854                 return perf_env__get_cpu_aggr_by_cpu;
1855         case AGGR_THREAD:
1856         case AGGR_UNSET:
1857         case AGGR_MAX:
1858         default:
1859                 return NULL;
1860         }
1861 }
1862
1863 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1864 {
1865         switch (aggr_mode) {
1866         case AGGR_SOCKET:
1867                 return perf_stat__get_socket_file;
1868         case AGGR_DIE:
1869                 return perf_stat__get_die_file;
1870         case AGGR_CACHE:
1871                 return perf_stat__get_cache_file;
1872         case AGGR_CORE:
1873                 return perf_stat__get_core_file;
1874         case AGGR_NODE:
1875                 return perf_stat__get_node_file;
1876         case AGGR_GLOBAL:
1877                 return perf_stat__get_global_file;
1878         case AGGR_NONE:
1879                 return perf_stat__get_cpu_file;
1880         case AGGR_THREAD:
1881         case AGGR_UNSET:
1882         case AGGR_MAX:
1883         default:
1884                 return NULL;
1885         }
1886 }
1887
1888 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1889 {
1890         struct perf_env *env = &st->session->header.env;
1891         aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1892         bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1893
1894         if (stat_config.aggr_mode == AGGR_THREAD) {
1895                 int nr = perf_thread_map__nr(evsel_list->core.threads);
1896
1897                 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1898                 if (stat_config.aggr_map == NULL)
1899                         return -ENOMEM;
1900
1901                 for (int s = 0; s < nr; s++) {
1902                         struct aggr_cpu_id id = aggr_cpu_id__empty();
1903
1904                         id.thread_idx = s;
1905                         stat_config.aggr_map->map[s] = id;
1906                 }
1907                 return 0;
1908         }
1909
1910         if (!get_id)
1911                 return 0;
1912
1913         stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1914                                                  get_id, env, needs_sort);
1915         if (!stat_config.aggr_map) {
1916                 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1917                 return -1;
1918         }
1919         stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1920         return 0;
1921 }
1922
1923 /*
1924  * Add default attributes, if there were no attributes specified or
1925  * if -d/--detailed, -d -d or -d -d -d is used:
1926  */
1927 static int add_default_attributes(void)
1928 {
1929         struct perf_event_attr default_attrs0[] = {
1930
1931   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1932   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1933   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1934   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1935
1936   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1937 };
1938         struct perf_event_attr frontend_attrs[] = {
1939   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1940 };
1941         struct perf_event_attr backend_attrs[] = {
1942   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1943 };
1944         struct perf_event_attr default_attrs1[] = {
1945   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1946   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1947   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1948
1949 };
1950
1951 /*
1952  * Detailed stats (-d), covering the L1 and last level data caches:
1953  */
1954         struct perf_event_attr detailed_attrs[] = {
1955
1956   { .type = PERF_TYPE_HW_CACHE,
1957     .config =
1958          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1959         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1960         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1961
1962   { .type = PERF_TYPE_HW_CACHE,
1963     .config =
1964          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1965         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1966         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1967
1968   { .type = PERF_TYPE_HW_CACHE,
1969     .config =
1970          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1971         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1972         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1973
1974   { .type = PERF_TYPE_HW_CACHE,
1975     .config =
1976          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1977         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1978         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1979 };
1980
1981 /*
1982  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1983  */
1984         struct perf_event_attr very_detailed_attrs[] = {
1985
1986   { .type = PERF_TYPE_HW_CACHE,
1987     .config =
1988          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1989         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1990         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1991
1992   { .type = PERF_TYPE_HW_CACHE,
1993     .config =
1994          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1995         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1996         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1997
1998   { .type = PERF_TYPE_HW_CACHE,
1999     .config =
2000          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2001         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2002         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2003
2004   { .type = PERF_TYPE_HW_CACHE,
2005     .config =
2006          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2007         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2008         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2009
2010   { .type = PERF_TYPE_HW_CACHE,
2011     .config =
2012          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2013         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2014         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2015
2016   { .type = PERF_TYPE_HW_CACHE,
2017     .config =
2018          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2019         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2020         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2021
2022 };
2023
2024 /*
2025  * Very, very detailed stats (-d -d -d), adding prefetch events:
2026  */
2027         struct perf_event_attr very_very_detailed_attrs[] = {
2028
2029   { .type = PERF_TYPE_HW_CACHE,
2030     .config =
2031          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2032         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2033         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2034
2035   { .type = PERF_TYPE_HW_CACHE,
2036     .config =
2037          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2038         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2039         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2040 };
2041
2042         struct perf_event_attr default_null_attrs[] = {};
2043         const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2044
2045         /* Set attrs if no event is selected and !null_run: */
2046         if (stat_config.null_run)
2047                 return 0;
2048
2049         if (transaction_run) {
2050                 /* Handle -T as -M transaction. Once platform specific metrics
2051                  * support has been added to the json files, all architectures
2052                  * will use this approach. To determine transaction support
2053                  * on an architecture test for such a metric name.
2054                  */
2055                 if (!metricgroup__has_metric(pmu, "transaction")) {
2056                         pr_err("Missing transaction metrics\n");
2057                         return -1;
2058                 }
2059                 return metricgroup__parse_groups(evsel_list, pmu, "transaction",
2060                                                 stat_config.metric_no_group,
2061                                                 stat_config.metric_no_merge,
2062                                                 stat_config.metric_no_threshold,
2063                                                 stat_config.user_requested_cpu_list,
2064                                                 stat_config.system_wide,
2065                                                 &stat_config.metric_events);
2066         }
2067
2068         if (smi_cost) {
2069                 int smi;
2070
2071                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2072                         pr_err("freeze_on_smi is not supported.\n");
2073                         return -1;
2074                 }
2075
2076                 if (!smi) {
2077                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2078                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2079                                 return -1;
2080                         }
2081                         smi_reset = true;
2082                 }
2083
2084                 if (!metricgroup__has_metric(pmu, "smi")) {
2085                         pr_err("Missing smi metrics\n");
2086                         return -1;
2087                 }
2088
2089                 if (!force_metric_only)
2090                         stat_config.metric_only = true;
2091
2092                 return metricgroup__parse_groups(evsel_list, pmu, "smi",
2093                                                 stat_config.metric_no_group,
2094                                                 stat_config.metric_no_merge,
2095                                                 stat_config.metric_no_threshold,
2096                                                 stat_config.user_requested_cpu_list,
2097                                                 stat_config.system_wide,
2098                                                 &stat_config.metric_events);
2099         }
2100
2101         if (topdown_run) {
2102                 unsigned int max_level = metricgroups__topdown_max_level();
2103                 char str[] = "TopdownL1";
2104
2105                 if (!force_metric_only)
2106                         stat_config.metric_only = true;
2107
2108                 if (!max_level) {
2109                         pr_err("Topdown requested but the topdown metric groups aren't present.\n"
2110                                 "(See perf list the metric groups have names like TopdownL1)\n");
2111                         return -1;
2112                 }
2113                 if (stat_config.topdown_level > max_level) {
2114                         pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
2115                         return -1;
2116                 } else if (!stat_config.topdown_level)
2117                         stat_config.topdown_level = 1;
2118
2119                 if (!stat_config.interval && !stat_config.metric_only) {
2120                         fprintf(stat_config.output,
2121                                 "Topdown accuracy may decrease when measuring long periods.\n"
2122                                 "Please print the result regularly, e.g. -I1000\n");
2123                 }
2124                 str[8] = stat_config.topdown_level + '0';
2125                 if (metricgroup__parse_groups(evsel_list,
2126                                                 pmu, str,
2127                                                 /*metric_no_group=*/false,
2128                                                 /*metric_no_merge=*/false,
2129                                                 /*metric_no_threshold=*/true,
2130                                                 stat_config.user_requested_cpu_list,
2131                                                 stat_config.system_wide,
2132                                                 &stat_config.metric_events) < 0)
2133                         return -1;
2134         }
2135
2136         if (!stat_config.topdown_level)
2137                 stat_config.topdown_level = 1;
2138
2139         if (!evsel_list->core.nr_entries) {
2140                 /* No events so add defaults. */
2141                 if (target__has_cpu(&target))
2142                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2143
2144                 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2145                         return -1;
2146                 if (perf_pmus__have_event("cpu", "stalled-cycles-frontend")) {
2147                         if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
2148                                 return -1;
2149                 }
2150                 if (perf_pmus__have_event("cpu", "stalled-cycles-backend")) {
2151                         if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2152                                 return -1;
2153                 }
2154                 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2155                         return -1;
2156                 /*
2157                  * Add TopdownL1 metrics if they exist. To minimize
2158                  * multiplexing, don't request threshold computation.
2159                  */
2160                 if (metricgroup__has_metric(pmu, "Default")) {
2161                         struct evlist *metric_evlist = evlist__new();
2162                         struct evsel *metric_evsel;
2163
2164                         if (!metric_evlist)
2165                                 return -1;
2166
2167                         if (metricgroup__parse_groups(metric_evlist, pmu, "Default",
2168                                                         /*metric_no_group=*/false,
2169                                                         /*metric_no_merge=*/false,
2170                                                         /*metric_no_threshold=*/true,
2171                                                         stat_config.user_requested_cpu_list,
2172                                                         stat_config.system_wide,
2173                                                         &stat_config.metric_events) < 0)
2174                                 return -1;
2175
2176                         evlist__for_each_entry(metric_evlist, metric_evsel) {
2177                                 metric_evsel->skippable = true;
2178                                 metric_evsel->default_metricgroup = true;
2179                         }
2180                         evlist__splice_list_tail(evsel_list, &metric_evlist->core.entries);
2181                         evlist__delete(metric_evlist);
2182                 }
2183
2184                 /* Platform specific attrs */
2185                 if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2186                         return -1;
2187         }
2188
2189         /* Detailed events get appended to the event list: */
2190
2191         if (detailed_run <  1)
2192                 return 0;
2193
2194         /* Append detailed run extra attributes: */
2195         if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2196                 return -1;
2197
2198         if (detailed_run < 2)
2199                 return 0;
2200
2201         /* Append very detailed run extra attributes: */
2202         if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2203                 return -1;
2204
2205         if (detailed_run < 3)
2206                 return 0;
2207
2208         /* Append very, very detailed run extra attributes: */
2209         return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2210 }
2211
2212 static const char * const stat_record_usage[] = {
2213         "perf stat record [<options>]",
2214         NULL,
2215 };
2216
2217 static void init_features(struct perf_session *session)
2218 {
2219         int feat;
2220
2221         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2222                 perf_header__set_feat(&session->header, feat);
2223
2224         perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2225         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2226         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2227         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2228         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2229 }
2230
2231 static int __cmd_record(int argc, const char **argv)
2232 {
2233         struct perf_session *session;
2234         struct perf_data *data = &perf_stat.data;
2235
2236         argc = parse_options(argc, argv, stat_options, stat_record_usage,
2237                              PARSE_OPT_STOP_AT_NON_OPTION);
2238
2239         if (output_name)
2240                 data->path = output_name;
2241
2242         if (stat_config.run_count != 1 || forever) {
2243                 pr_err("Cannot use -r option with perf stat record.\n");
2244                 return -1;
2245         }
2246
2247         session = perf_session__new(data, NULL);
2248         if (IS_ERR(session)) {
2249                 pr_err("Perf session creation failed\n");
2250                 return PTR_ERR(session);
2251         }
2252
2253         init_features(session);
2254
2255         session->evlist   = evsel_list;
2256         perf_stat.session = session;
2257         perf_stat.record  = true;
2258         return argc;
2259 }
2260
2261 static int process_stat_round_event(struct perf_session *session,
2262                                     union perf_event *event)
2263 {
2264         struct perf_record_stat_round *stat_round = &event->stat_round;
2265         struct timespec tsh, *ts = NULL;
2266         const char **argv = session->header.env.cmdline_argv;
2267         int argc = session->header.env.nr_cmdline;
2268
2269         process_counters();
2270
2271         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2272                 update_stats(&walltime_nsecs_stats, stat_round->time);
2273
2274         if (stat_config.interval && stat_round->time) {
2275                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2276                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2277                 ts = &tsh;
2278         }
2279
2280         print_counters(ts, argc, argv);
2281         return 0;
2282 }
2283
2284 static
2285 int process_stat_config_event(struct perf_session *session,
2286                               union perf_event *event)
2287 {
2288         struct perf_tool *tool = session->tool;
2289         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2290
2291         perf_event__read_stat_config(&stat_config, &event->stat_config);
2292
2293         if (perf_cpu_map__has_any_cpu_or_is_empty(st->cpus)) {
2294                 if (st->aggr_mode != AGGR_UNSET)
2295                         pr_warning("warning: processing task data, aggregation mode not set\n");
2296         } else if (st->aggr_mode != AGGR_UNSET) {
2297                 stat_config.aggr_mode = st->aggr_mode;
2298         }
2299
2300         if (perf_stat.data.is_pipe)
2301                 perf_stat_init_aggr_mode();
2302         else
2303                 perf_stat_init_aggr_mode_file(st);
2304
2305         if (stat_config.aggr_map) {
2306                 int nr_aggr = stat_config.aggr_map->nr;
2307
2308                 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2309                         pr_err("cannot allocate aggr counts\n");
2310                         return -1;
2311                 }
2312         }
2313         return 0;
2314 }
2315
2316 static int set_maps(struct perf_stat *st)
2317 {
2318         if (!st->cpus || !st->threads)
2319                 return 0;
2320
2321         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2322                 return -EINVAL;
2323
2324         perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2325
2326         if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2327                 return -ENOMEM;
2328
2329         st->maps_allocated = true;
2330         return 0;
2331 }
2332
2333 static
2334 int process_thread_map_event(struct perf_session *session,
2335                              union perf_event *event)
2336 {
2337         struct perf_tool *tool = session->tool;
2338         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2339
2340         if (st->threads) {
2341                 pr_warning("Extra thread map event, ignoring.\n");
2342                 return 0;
2343         }
2344
2345         st->threads = thread_map__new_event(&event->thread_map);
2346         if (!st->threads)
2347                 return -ENOMEM;
2348
2349         return set_maps(st);
2350 }
2351
2352 static
2353 int process_cpu_map_event(struct perf_session *session,
2354                           union perf_event *event)
2355 {
2356         struct perf_tool *tool = session->tool;
2357         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2358         struct perf_cpu_map *cpus;
2359
2360         if (st->cpus) {
2361                 pr_warning("Extra cpu map event, ignoring.\n");
2362                 return 0;
2363         }
2364
2365         cpus = cpu_map__new_data(&event->cpu_map.data);
2366         if (!cpus)
2367                 return -ENOMEM;
2368
2369         st->cpus = cpus;
2370         return set_maps(st);
2371 }
2372
2373 static const char * const stat_report_usage[] = {
2374         "perf stat report [<options>]",
2375         NULL,
2376 };
2377
2378 static struct perf_stat perf_stat = {
2379         .tool = {
2380                 .attr           = perf_event__process_attr,
2381                 .event_update   = perf_event__process_event_update,
2382                 .thread_map     = process_thread_map_event,
2383                 .cpu_map        = process_cpu_map_event,
2384                 .stat_config    = process_stat_config_event,
2385                 .stat           = perf_event__process_stat_event,
2386                 .stat_round     = process_stat_round_event,
2387         },
2388         .aggr_mode      = AGGR_UNSET,
2389         .aggr_level     = 0,
2390 };
2391
2392 static int __cmd_report(int argc, const char **argv)
2393 {
2394         struct perf_session *session;
2395         const struct option options[] = {
2396         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2397         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2398                      "aggregate counts per processor socket", AGGR_SOCKET),
2399         OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2400                      "aggregate counts per processor die", AGGR_DIE),
2401         OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
2402                             "cache level",
2403                             "aggregate count at this cache level (Default: LLC)",
2404                             parse_cache_level),
2405         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2406                      "aggregate counts per physical processor core", AGGR_CORE),
2407         OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2408                      "aggregate counts per numa node", AGGR_NODE),
2409         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2410                      "disable CPU count aggregation", AGGR_NONE),
2411         OPT_END()
2412         };
2413         struct stat st;
2414         int ret;
2415
2416         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2417
2418         if (!input_name || !strlen(input_name)) {
2419                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2420                         input_name = "-";
2421                 else
2422                         input_name = "perf.data";
2423         }
2424
2425         perf_stat.data.path = input_name;
2426         perf_stat.data.mode = PERF_DATA_MODE_READ;
2427
2428         session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2429         if (IS_ERR(session))
2430                 return PTR_ERR(session);
2431
2432         perf_stat.session  = session;
2433         stat_config.output = stderr;
2434         evlist__delete(evsel_list);
2435         evsel_list         = session->evlist;
2436
2437         ret = perf_session__process_events(session);
2438         if (ret)
2439                 return ret;
2440
2441         perf_session__delete(session);
2442         return 0;
2443 }
2444
2445 static void setup_system_wide(int forks)
2446 {
2447         /*
2448          * Make system wide (-a) the default target if
2449          * no target was specified and one of following
2450          * conditions is met:
2451          *
2452          *   - there's no workload specified
2453          *   - there is workload specified but all requested
2454          *     events are system wide events
2455          */
2456         if (!target__none(&target))
2457                 return;
2458
2459         if (!forks)
2460                 target.system_wide = true;
2461         else {
2462                 struct evsel *counter;
2463
2464                 evlist__for_each_entry(evsel_list, counter) {
2465                         if (!counter->core.requires_cpu &&
2466                             !evsel__name_is(counter, "duration_time")) {
2467                                 return;
2468                         }
2469                 }
2470
2471                 if (evsel_list->core.nr_entries)
2472                         target.system_wide = true;
2473         }
2474 }
2475
2476 int cmd_stat(int argc, const char **argv)
2477 {
2478         const char * const stat_usage[] = {
2479                 "perf stat [<options>] [<command>]",
2480                 NULL
2481         };
2482         int status = -EINVAL, run_idx, err;
2483         const char *mode;
2484         FILE *output = stderr;
2485         unsigned int interval, timeout;
2486         const char * const stat_subcommands[] = { "record", "report" };
2487         char errbuf[BUFSIZ];
2488
2489         setlocale(LC_ALL, "");
2490
2491         evsel_list = evlist__new();
2492         if (evsel_list == NULL)
2493                 return -ENOMEM;
2494
2495         parse_events__shrink_config_terms();
2496
2497         /* String-parsing callback-based options would segfault when negated */
2498         set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2499         set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2500         set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2501
2502         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2503                                         (const char **) stat_usage,
2504                                         PARSE_OPT_STOP_AT_NON_OPTION);
2505
2506         if (stat_config.csv_sep) {
2507                 stat_config.csv_output = true;
2508                 if (!strcmp(stat_config.csv_sep, "\\t"))
2509                         stat_config.csv_sep = "\t";
2510         } else
2511                 stat_config.csv_sep = DEFAULT_SEPARATOR;
2512
2513         if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2514                 argc = __cmd_record(argc, argv);
2515                 if (argc < 0)
2516                         return -1;
2517         } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2518                 return __cmd_report(argc, argv);
2519
2520         interval = stat_config.interval;
2521         timeout = stat_config.timeout;
2522
2523         /*
2524          * For record command the -o is already taken care of.
2525          */
2526         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2527                 output = NULL;
2528
2529         if (output_name && output_fd) {
2530                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2531                 parse_options_usage(stat_usage, stat_options, "o", 1);
2532                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2533                 goto out;
2534         }
2535
2536         if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2537                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2538                 goto out;
2539         }
2540
2541         if (stat_config.metric_only && stat_config.run_count > 1) {
2542                 fprintf(stderr, "--metric-only is not supported with -r\n");
2543                 goto out;
2544         }
2545
2546         if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2547                 fprintf(stderr, "--table is only supported with -r\n");
2548                 parse_options_usage(stat_usage, stat_options, "r", 1);
2549                 parse_options_usage(NULL, stat_options, "table", 0);
2550                 goto out;
2551         }
2552
2553         if (output_fd < 0) {
2554                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2555                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2556                 goto out;
2557         }
2558
2559         if (!output && !quiet) {
2560                 struct timespec tm;
2561                 mode = append_file ? "a" : "w";
2562
2563                 output = fopen(output_name, mode);
2564                 if (!output) {
2565                         perror("failed to create output file");
2566                         return -1;
2567                 }
2568                 if (!stat_config.json_output) {
2569                         clock_gettime(CLOCK_REALTIME, &tm);
2570                         fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2571                 }
2572         } else if (output_fd > 0) {
2573                 mode = append_file ? "a" : "w";
2574                 output = fdopen(output_fd, mode);
2575                 if (!output) {
2576                         perror("Failed opening logfd");
2577                         return -errno;
2578                 }
2579         }
2580
2581         if (stat_config.interval_clear && !isatty(fileno(output))) {
2582                 fprintf(stderr, "--interval-clear does not work with output\n");
2583                 parse_options_usage(stat_usage, stat_options, "o", 1);
2584                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2585                 parse_options_usage(NULL, stat_options, "interval-clear", 0);
2586                 return -1;
2587         }
2588
2589         stat_config.output = output;
2590
2591         /*
2592          * let the spreadsheet do the pretty-printing
2593          */
2594         if (stat_config.csv_output) {
2595                 /* User explicitly passed -B? */
2596                 if (big_num_opt == 1) {
2597                         fprintf(stderr, "-B option not supported with -x\n");
2598                         parse_options_usage(stat_usage, stat_options, "B", 1);
2599                         parse_options_usage(NULL, stat_options, "x", 1);
2600                         goto out;
2601                 } else /* Nope, so disable big number formatting */
2602                         stat_config.big_num = false;
2603         } else if (big_num_opt == 0) /* User passed --no-big-num */
2604                 stat_config.big_num = false;
2605
2606         err = target__validate(&target);
2607         if (err) {
2608                 target__strerror(&target, err, errbuf, BUFSIZ);
2609                 pr_warning("%s\n", errbuf);
2610         }
2611
2612         setup_system_wide(argc);
2613
2614         /*
2615          * Display user/system times only for single
2616          * run and when there's specified tracee.
2617          */
2618         if ((stat_config.run_count == 1) && target__none(&target))
2619                 stat_config.ru_display = true;
2620
2621         if (stat_config.run_count < 0) {
2622                 pr_err("Run count must be a positive number\n");
2623                 parse_options_usage(stat_usage, stat_options, "r", 1);
2624                 goto out;
2625         } else if (stat_config.run_count == 0) {
2626                 forever = true;
2627                 stat_config.run_count = 1;
2628         }
2629
2630         if (stat_config.walltime_run_table) {
2631                 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2632                 if (!stat_config.walltime_run) {
2633                         pr_err("failed to setup -r option");
2634                         goto out;
2635                 }
2636         }
2637
2638         if ((stat_config.aggr_mode == AGGR_THREAD) &&
2639                 !target__has_task(&target)) {
2640                 if (!target.system_wide || target.cpu_list) {
2641                         fprintf(stderr, "The --per-thread option is only "
2642                                 "available when monitoring via -p -t -a "
2643                                 "options or only --per-thread.\n");
2644                         parse_options_usage(NULL, stat_options, "p", 1);
2645                         parse_options_usage(NULL, stat_options, "t", 1);
2646                         goto out;
2647                 }
2648         }
2649
2650         /*
2651          * no_aggr, cgroup are for system-wide only
2652          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2653          */
2654         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2655               stat_config.aggr_mode != AGGR_THREAD) ||
2656              (nr_cgroups || stat_config.cgroup_list)) &&
2657             !target__has_cpu(&target)) {
2658                 fprintf(stderr, "both cgroup and no-aggregation "
2659                         "modes only available in system-wide mode\n");
2660
2661                 parse_options_usage(stat_usage, stat_options, "G", 1);
2662                 parse_options_usage(NULL, stat_options, "A", 1);
2663                 parse_options_usage(NULL, stat_options, "a", 1);
2664                 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2665                 goto out;
2666         }
2667
2668         if (stat_config.iostat_run) {
2669                 status = iostat_prepare(evsel_list, &stat_config);
2670                 if (status)
2671                         goto out;
2672                 if (iostat_mode == IOSTAT_LIST) {
2673                         iostat_list(evsel_list, &stat_config);
2674                         goto out;
2675                 } else if (verbose > 0)
2676                         iostat_list(evsel_list, &stat_config);
2677                 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2678                         target.system_wide = true;
2679         }
2680
2681         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2682                 target.per_thread = true;
2683
2684         stat_config.system_wide = target.system_wide;
2685         if (target.cpu_list) {
2686                 stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2687                 if (!stat_config.user_requested_cpu_list) {
2688                         status = -ENOMEM;
2689                         goto out;
2690                 }
2691         }
2692
2693         /*
2694          * Metric parsing needs to be delayed as metrics may optimize events
2695          * knowing the target is system-wide.
2696          */
2697         if (metrics) {
2698                 const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2699                 int ret = metricgroup__parse_groups(evsel_list, pmu, metrics,
2700                                                 stat_config.metric_no_group,
2701                                                 stat_config.metric_no_merge,
2702                                                 stat_config.metric_no_threshold,
2703                                                 stat_config.user_requested_cpu_list,
2704                                                 stat_config.system_wide,
2705                                                 &stat_config.metric_events);
2706
2707                 zfree(&metrics);
2708                 if (ret) {
2709                         status = ret;
2710                         goto out;
2711                 }
2712         }
2713
2714         if (add_default_attributes())
2715                 goto out;
2716
2717         if (stat_config.cgroup_list) {
2718                 if (nr_cgroups > 0) {
2719                         pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2720                         parse_options_usage(stat_usage, stat_options, "G", 1);
2721                         parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2722                         goto out;
2723                 }
2724
2725                 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2726                                           &stat_config.metric_events, true) < 0) {
2727                         parse_options_usage(stat_usage, stat_options,
2728                                             "for-each-cgroup", 0);
2729                         goto out;
2730                 }
2731         }
2732
2733         evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
2734
2735         if (evlist__create_maps(evsel_list, &target) < 0) {
2736                 if (target__has_task(&target)) {
2737                         pr_err("Problems finding threads of monitor\n");
2738                         parse_options_usage(stat_usage, stat_options, "p", 1);
2739                         parse_options_usage(NULL, stat_options, "t", 1);
2740                 } else if (target__has_cpu(&target)) {
2741                         perror("failed to parse CPUs map");
2742                         parse_options_usage(stat_usage, stat_options, "C", 1);
2743                         parse_options_usage(NULL, stat_options, "a", 1);
2744                 }
2745                 goto out;
2746         }
2747
2748         evlist__check_cpu_maps(evsel_list);
2749
2750         /*
2751          * Initialize thread_map with comm names,
2752          * so we could print it out on output.
2753          */
2754         if (stat_config.aggr_mode == AGGR_THREAD) {
2755                 thread_map__read_comms(evsel_list->core.threads);
2756         }
2757
2758         if (stat_config.aggr_mode == AGGR_NODE)
2759                 cpu__setup_cpunode_map();
2760
2761         if (stat_config.times && interval)
2762                 interval_count = true;
2763         else if (stat_config.times && !interval) {
2764                 pr_err("interval-count option should be used together with "
2765                                 "interval-print.\n");
2766                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2767                 parse_options_usage(stat_usage, stat_options, "I", 1);
2768                 goto out;
2769         }
2770
2771         if (timeout && timeout < 100) {
2772                 if (timeout < 10) {
2773                         pr_err("timeout must be >= 10ms.\n");
2774                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
2775                         goto out;
2776                 } else
2777                         pr_warning("timeout < 100ms. "
2778                                    "The overhead percentage could be high in some cases. "
2779                                    "Please proceed with caution.\n");
2780         }
2781         if (timeout && interval) {
2782                 pr_err("timeout option is not supported with interval-print.\n");
2783                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2784                 parse_options_usage(stat_usage, stat_options, "I", 1);
2785                 goto out;
2786         }
2787
2788         if (perf_stat_init_aggr_mode())
2789                 goto out;
2790
2791         if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2792                 goto out;
2793
2794         /*
2795          * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2796          * while avoiding that older tools show confusing messages.
2797          *
2798          * However for pipe sessions we need to keep it zero,
2799          * because script's perf_evsel__check_attr is triggered
2800          * by attr->sample_type != 0, and we can't run it on
2801          * stat sessions.
2802          */
2803         stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2804
2805         /*
2806          * We dont want to block the signals - that would cause
2807          * child tasks to inherit that and Ctrl-C would not work.
2808          * What we want is for Ctrl-C to work in the exec()-ed
2809          * task, but being ignored by perf stat itself:
2810          */
2811         atexit(sig_atexit);
2812         if (!forever)
2813                 signal(SIGINT,  skip_signal);
2814         signal(SIGCHLD, skip_signal);
2815         signal(SIGALRM, skip_signal);
2816         signal(SIGABRT, skip_signal);
2817
2818         if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2819                 goto out;
2820
2821         /* Enable ignoring missing threads when -p option is defined. */
2822         evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2823         status = 0;
2824         for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2825                 if (stat_config.run_count != 1 && verbose > 0)
2826                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2827                                 run_idx + 1);
2828
2829                 if (run_idx != 0)
2830                         evlist__reset_prev_raw_counts(evsel_list);
2831
2832                 status = run_perf_stat(argc, argv, run_idx);
2833                 if (forever && status != -1 && !interval) {
2834                         print_counters(NULL, argc, argv);
2835                         perf_stat__reset_stats();
2836                 }
2837         }
2838
2839         if (!forever && status != -1 && (!interval || stat_config.summary)) {
2840                 if (stat_config.run_count > 1)
2841                         evlist__copy_res_stats(&stat_config, evsel_list);
2842                 print_counters(NULL, argc, argv);
2843         }
2844
2845         evlist__finalize_ctlfd(evsel_list);
2846
2847         if (STAT_RECORD) {
2848                 /*
2849                  * We synthesize the kernel mmap record just so that older tools
2850                  * don't emit warnings about not being able to resolve symbols
2851                  * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2852                  * a saner message about no samples being in the perf.data file.
2853                  *
2854                  * This also serves to suppress a warning about f_header.data.size == 0
2855                  * in header.c at the moment 'perf stat record' gets introduced, which
2856                  * is not really needed once we start adding the stat specific PERF_RECORD_
2857                  * records, but the need to suppress the kptr_restrict messages in older
2858                  * tools remain  -acme
2859                  */
2860                 int fd = perf_data__fd(&perf_stat.data);
2861
2862                 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2863                                                          process_synthesized_event,
2864                                                          &perf_stat.session->machines.host);
2865                 if (err) {
2866                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2867                                    "older tools may produce warnings about this file\n.");
2868                 }
2869
2870                 if (!interval) {
2871                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2872                                 pr_err("failed to write stat round event\n");
2873                 }
2874
2875                 if (!perf_stat.data.is_pipe) {
2876                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2877                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2878                 }
2879
2880                 evlist__close(evsel_list);
2881                 perf_session__delete(perf_stat.session);
2882         }
2883
2884         perf_stat__exit_aggr_mode();
2885         evlist__free_stats(evsel_list);
2886 out:
2887         if (stat_config.iostat_run)
2888                 iostat_release(evsel_list);
2889
2890         zfree(&stat_config.walltime_run);
2891         zfree(&stat_config.user_requested_cpu_list);
2892
2893         if (smi_cost && smi_reset)
2894                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2895
2896         evlist__delete(evsel_list);
2897
2898         metricgroup__rblist_exit(&stat_config.metric_events);
2899         evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2900
2901         return status;
2902 }