tools/perf/builtin-stat.c

   1 /*
   2  * builtin-stat.c
   3  *
   4  * Builtin stat command: Give a precise performance counters summary
   5  * overview about any workload, CPU or specific PID.
   6  *
   7  * Sample output:
   8
   9    $ perf stat ./hackbench 10
  10
  11   Time: 0.118
  12
  13   Performance counter stats for './hackbench 10':
  14
  15        1708.761321 task-clock                #   11.037 CPUs utilized
  16             41,190 context-switches          #    0.024 M/sec
  17              6,735 CPU-migrations            #    0.004 M/sec
  18             17,318 page-faults               #    0.010 M/sec
  19      5,205,202,243 cycles                    #    3.046 GHz
  20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  22      2,603,501,247 instructions              #    0.50  insns per cycle
  23                                              #    1.48  stalled cycles per insn
  24        484,357,498 branches                  #  283.455 M/sec
  25          6,388,934 branch-misses             #    1.32% of all branches
  26
  27         0.154822978  seconds time elapsed
  28
  29  *
  30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  31  *
  32  * Improvements and fixes by:
  33  *
  34  *   Arjan van de Ven <arjan@linux.intel.com>
  35  *   Yanmin Zhang <yanmin.zhang@intel.com>
  36  *   Wu Fengguang <fengguang.wu@intel.com>
  37  *   Mike Galbraith <efault@gmx.de>
  38  *   Paul Mackerras <paulus@samba.org>
  39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  40  *
  41  * Released under the GPL v2. (and only v2, not any later version)
  42  */
  43
  44 #include "perf.h"
  45 #include "builtin.h"
  46 #include "util/cgroup.h"
  47 #include "util/util.h"
  48 #include "util/parse-options.h"
  49 #include "util/parse-events.h"
  50 #include "util/pmu.h"
  51 #include "util/event.h"
  52 #include "util/evlist.h"
  53 #include "util/evsel.h"
  54 #include "util/debug.h"
  55 #include "util/color.h"
  56 #include "util/stat.h"
  57 #include "util/header.h"
  58 #include "util/cpumap.h"
  59 #include "util/thread.h"
  60 #include "util/thread_map.h"
  61 #include "util/counts.h"
  62
  63 #include <stdlib.h>
  64 #include <sys/prctl.h>
  65 #include <locale.h>
  66
  67 #define DEFAULT_SEPARATOR       " "
  68 #define CNTR_NOT_SUPPORTED      "<not supported>"
  69 #define CNTR_NOT_COUNTED        "<not counted>"
  70
  71 static void print_counters(struct timespec *ts, int argc, const char **argv);
  72
  73 /* Default events used for perf stat -T */
  74 static const char *transaction_attrs = {
  75         "task-clock,"
  76         "{"
  77         "instructions,"
  78         "cycles,"
  79         "cpu/cycles-t/,"
  80         "cpu/tx-start/,"
  81         "cpu/el-start/,"
  82         "cpu/cycles-ct/"
  83         "}"
  84 };
  85
  86 /* More limited version when the CPU does not have all events. */
  87 static const char * transaction_limited_attrs = {
  88         "task-clock,"
  89         "{"
  90         "instructions,"
  91         "cycles,"
  92         "cpu/cycles-t/,"
  93         "cpu/tx-start/"
  94         "}"
  95 };
  96
  97 static struct perf_evlist       *evsel_list;
  98
  99 static struct target target = {
 100         .uid    = UINT_MAX,
 101 };
 102
 103 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
 104
 105 static int                      run_count                       =  1;
 106 static bool                     no_inherit                      = false;
 107 static volatile pid_t           child_pid                       = -1;
 108 static bool                     null_run                        =  false;
 109 static int                      detailed_run                    =  0;
 110 static bool                     transaction_run;
 111 static bool                     big_num                         =  true;
 112 static int                      big_num_opt                     =  -1;
 113 static const char               *csv_sep                        = NULL;
 114 static bool                     csv_output                      = false;
 115 static bool                     group                           = false;
 116 static const char               *pre_cmd                        = NULL;
 117 static const char               *post_cmd                       = NULL;
 118 static bool                     sync_run                        = false;
 119 static unsigned int             initial_delay                   = 0;
 120 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 121 static bool                     forever                         = false;
 122 static struct timespec          ref_time;
 123 static struct cpu_map           *aggr_map;
 124 static aggr_get_id_t            aggr_get_id;
 125
 126 static volatile int done = 0;
 127
 128 static struct perf_stat_config stat_config = {
 129         .aggr_mode      = AGGR_GLOBAL,
 130         .scale          = true,
 131 };
 132
 133 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 134                                  struct timespec *b)
 135 {
 136         r->tv_sec = a->tv_sec - b->tv_sec;
 137         if (a->tv_nsec < b->tv_nsec) {
 138                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 139                 r->tv_sec--;
 140         } else {
 141                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 142         }
 143 }
 144
 145 static void perf_stat__reset_stats(void)
 146 {
 147         perf_evlist__reset_stats(evsel_list);
 148         perf_stat__reset_shadow_stats();
 149 }
 150
 151 static int create_perf_stat_counter(struct perf_evsel *evsel)
 152 {
 153         struct perf_event_attr *attr = &evsel->attr;
 154
 155         if (stat_config.scale)
 156                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 157                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 158
 159         attr->inherit = !no_inherit;
 160
 161         if (target__has_cpu(&target))
 162                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 163
 164         if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
 165                 attr->disabled = 1;
 166                 if (!initial_delay)
 167                         attr->enable_on_exec = 1;
 168         }
 169
 170         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 171 }
 172
 173 /*
 174  * Does the counter have nsecs as a unit?
 175  */
 176 static inline int nsec_counter(struct perf_evsel *evsel)
 177 {
 178         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 179             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 180                 return 1;
 181
 182         return 0;
 183 }
 184
 185 /*
 186  * Read out the results of a single counter:
 187  * do not aggregate counts across CPUs in system-wide mode
 188  */
 189 static int read_counter(struct perf_evsel *counter)
 190 {
 191         int nthreads = thread_map__nr(evsel_list->threads);
 192         int ncpus = perf_evsel__nr_cpus(counter);
 193         int cpu, thread;
 194
 195         if (!counter->supported)
 196                 return -ENOENT;
 197
 198         if (counter->system_wide)
 199                 nthreads = 1;
 200
 201         for (thread = 0; thread < nthreads; thread++) {
 202                 for (cpu = 0; cpu < ncpus; cpu++) {
 203                         struct perf_counts_values *count;
 204
 205                         count = perf_counts(counter->counts, cpu, thread);
 206                         if (perf_evsel__read(counter, cpu, thread, count))
 207                                 return -1;
 208                 }
 209         }
 210
 211         return 0;
 212 }
 213
 214 static void read_counters(bool close_counters)
 215 {
 216         struct perf_evsel *counter;
 217
 218         evlist__for_each(evsel_list, counter) {
 219                 if (read_counter(counter))
 220                         pr_debug("failed to read counter %s\n", counter->name);
 221
 222                 if (perf_stat_process_counter(&stat_config, counter))
 223                         pr_warning("failed to process counter %s\n", counter->name);
 224
 225                 if (close_counters) {
 226                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 227                                              thread_map__nr(evsel_list->threads));
 228                 }
 229         }
 230 }
 231
 232 static void process_interval(void)
 233 {
 234         struct timespec ts, rs;
 235
 236         read_counters(false);
 237
 238         clock_gettime(CLOCK_MONOTONIC, &ts);
 239         diff_timespec(&rs, &ts, &ref_time);
 240
 241         print_counters(&rs, 0, NULL);
 242 }
 243
 244 static void handle_initial_delay(void)
 245 {
 246         struct perf_evsel *counter;
 247
 248         if (initial_delay) {
 249                 const int ncpus = cpu_map__nr(evsel_list->cpus),
 250                         nthreads = thread_map__nr(evsel_list->threads);
 251
 252                 usleep(initial_delay * 1000);
 253                 evlist__for_each(evsel_list, counter)
 254                         perf_evsel__enable(counter, ncpus, nthreads);
 255         }
 256 }
 257
 258 static volatile int workload_exec_errno;
 259
 260 /*
 261  * perf_evlist__prepare_workload will send a SIGUSR1
 262  * if the fork fails, since we asked by setting its
 263  * want_signal to true.
 264  */
 265 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 266                                         void *ucontext __maybe_unused)
 267 {
 268         workload_exec_errno = info->si_value.sival_int;
 269 }
 270
 271 static int __run_perf_stat(int argc, const char **argv)
 272 {
 273         int interval = stat_config.interval;
 274         char msg[512];
 275         unsigned long long t0, t1;
 276         struct perf_evsel *counter;
 277         struct timespec ts;
 278         size_t l;
 279         int status = 0;
 280         const bool forks = (argc > 0);
 281
 282         if (interval) {
 283                 ts.tv_sec  = interval / 1000;
 284                 ts.tv_nsec = (interval % 1000) * 1000000;
 285         } else {
 286                 ts.tv_sec  = 1;
 287                 ts.tv_nsec = 0;
 288         }
 289
 290         if (forks) {
 291                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
 292                                                   workload_exec_failed_signal) < 0) {
 293                         perror("failed to prepare workload");
 294                         return -1;
 295                 }
 296                 child_pid = evsel_list->workload.pid;
 297         }
 298
 299         if (group)
 300                 perf_evlist__set_leader(evsel_list);
 301
 302         evlist__for_each(evsel_list, counter) {
 303                 if (create_perf_stat_counter(counter) < 0) {
 304                         /*
 305                          * PPC returns ENXIO for HW counters until 2.6.37
 306                          * (behavior changed with commit b0a873e).
 307                          */
 308                         if (errno == EINVAL || errno == ENOSYS ||
 309                             errno == ENOENT || errno == EOPNOTSUPP ||
 310                             errno == ENXIO) {
 311                                 if (verbose)
 312                                         ui__warning("%s event is not supported by the kernel.\n",
 313                                                     perf_evsel__name(counter));
 314                                 counter->supported = false;
 315
 316                                 if ((counter->leader != counter) ||
 317                                     !(counter->leader->nr_members > 1))
 318                                         continue;
 319                         }
 320
 321                         perf_evsel__open_strerror(counter, &target,
 322                                                   errno, msg, sizeof(msg));
 323                         ui__error("%s\n", msg);
 324
 325                         if (child_pid != -1)
 326                                 kill(child_pid, SIGTERM);
 327
 328                         return -1;
 329                 }
 330                 counter->supported = true;
 331
 332                 l = strlen(counter->unit);
 333                 if (l > unit_width)
 334                         unit_width = l;
 335         }
 336
 337         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 338                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 339                         counter->filter, perf_evsel__name(counter), errno,
 340                         strerror_r(errno, msg, sizeof(msg)));
 341                 return -1;
 342         }
 343
 344         /*
 345          * Enable counters and exec the command:
 346          */
 347         t0 = rdclock();
 348         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 349
 350         if (forks) {
 351                 perf_evlist__start_workload(evsel_list);
 352                 handle_initial_delay();
 353
 354                 if (interval) {
 355                         while (!waitpid(child_pid, &status, WNOHANG)) {
 356                                 nanosleep(&ts, NULL);
 357                                 process_interval();
 358                         }
 359                 }
 360                 wait(&status);
 361
 362                 if (workload_exec_errno) {
 363                         const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 364                         pr_err("Workload failed: %s\n", emsg);
 365                         return -1;
 366                 }
 367
 368                 if (WIFSIGNALED(status))
 369                         psignal(WTERMSIG(status), argv[0]);
 370         } else {
 371                 handle_initial_delay();
 372                 while (!done) {
 373                         nanosleep(&ts, NULL);
 374                         if (interval)
 375                                 process_interval();
 376                 }
 377         }
 378
 379         t1 = rdclock();
 380
 381         update_stats(&walltime_nsecs_stats, t1 - t0);
 382
 383         read_counters(true);
 384
 385         return WEXITSTATUS(status);
 386 }
 387
 388 static int run_perf_stat(int argc, const char **argv)
 389 {
 390         int ret;
 391
 392         if (pre_cmd) {
 393                 ret = system(pre_cmd);
 394                 if (ret)
 395                         return ret;
 396         }
 397
 398         if (sync_run)
 399                 sync();
 400
 401         ret = __run_perf_stat(argc, argv);
 402         if (ret)
 403                 return ret;
 404
 405         if (post_cmd) {
 406                 ret = system(post_cmd);
 407                 if (ret)
 408                         return ret;
 409         }
 410
 411         return ret;
 412 }
 413
 414 static void print_running(u64 run, u64 ena)
 415 {
 416         if (csv_output) {
 417                 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
 418                                         csv_sep,
 419                                         run,
 420                                         csv_sep,
 421                                         ena ? 100.0 * run / ena : 100.0);
 422         } else if (run != ena) {
 423                 fprintf(stat_config.output, "  (%.2f%%)", 100.0 * run / ena);
 424         }
 425 }
 426
 427 static void print_noise_pct(double total, double avg)
 428 {
 429         double pct = rel_stddev_stats(total, avg);
 430
 431         if (csv_output)
 432                 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
 433         else if (pct)
 434                 fprintf(stat_config.output, "  ( +-%6.2f%% )", pct);
 435 }
 436
 437 static void print_noise(struct perf_evsel *evsel, double avg)
 438 {
 439         struct perf_stat_evsel *ps;
 440
 441         if (run_count == 1)
 442                 return;
 443
 444         ps = evsel->priv;
 445         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 446 }
 447
 448 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 449 {
 450         switch (stat_config.aggr_mode) {
 451         case AGGR_CORE:
 452                 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
 453                         cpu_map__id_to_socket(id),
 454                         csv_output ? 0 : -8,
 455                         cpu_map__id_to_cpu(id),
 456                         csv_sep,
 457                         csv_output ? 0 : 4,
 458                         nr,
 459                         csv_sep);
 460                 break;
 461         case AGGR_SOCKET:
 462                 fprintf(stat_config.output, "S%*d%s%*d%s",
 463                         csv_output ? 0 : -5,
 464                         id,
 465                         csv_sep,
 466                         csv_output ? 0 : 4,
 467                         nr,
 468                         csv_sep);
 469                         break;
 470         case AGGR_NONE:
 471                 fprintf(stat_config.output, "CPU%*d%s",
 472                         csv_output ? 0 : -4,
 473                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 474                 break;
 475         case AGGR_THREAD:
 476                 fprintf(stat_config.output, "%*s-%*d%s",
 477                         csv_output ? 0 : 16,
 478                         thread_map__comm(evsel->threads, id),
 479                         csv_output ? 0 : -8,
 480                         thread_map__pid(evsel->threads, id),
 481                         csv_sep);
 482                 break;
 483         case AGGR_GLOBAL:
 484         case AGGR_UNSET:
 485         default:
 486                 break;
 487         }
 488 }
 489
 490 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 491 {
 492         FILE *output = stat_config.output;
 493         double msecs = avg / 1e6;
 494         const char *fmt_v, *fmt_n;
 495         char name[25];
 496
 497         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
 498         fmt_n = csv_output ? "%s" : "%-25s";
 499
 500         aggr_printout(evsel, id, nr);
 501
 502         scnprintf(name, sizeof(name), "%s%s",
 503                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
 504
 505         fprintf(output, fmt_v, msecs, csv_sep);
 506
 507         if (csv_output)
 508                 fprintf(output, "%s%s", evsel->unit, csv_sep);
 509         else
 510                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
 511
 512         fprintf(output, fmt_n, name);
 513
 514         if (evsel->cgrp)
 515                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 516
 517         if (csv_output || stat_config.interval)
 518                 return;
 519
 520         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 521                 fprintf(output, " # %8.3f CPUs utilized          ",
 522                         avg / avg_stats(&walltime_nsecs_stats));
 523         else
 524                 fprintf(output, "                                   ");
 525 }
 526
 527 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 528 {
 529         FILE *output = stat_config.output;
 530         double sc =  evsel->scale;
 531         const char *fmt;
 532         int cpu = cpu_map__id_to_cpu(id);
 533
 534         if (csv_output) {
 535                 fmt = sc != 1.0 ?  "%.2f%s" : "%.0f%s";
 536         } else {
 537                 if (big_num)
 538                         fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
 539                 else
 540                         fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
 541         }
 542
 543         aggr_printout(evsel, id, nr);
 544
 545         if (stat_config.aggr_mode == AGGR_GLOBAL)
 546                 cpu = 0;
 547
 548         fprintf(output, fmt, avg, csv_sep);
 549
 550         if (evsel->unit)
 551                 fprintf(output, "%-*s%s",
 552                         csv_output ? 0 : unit_width,
 553                         evsel->unit, csv_sep);
 554
 555         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
 556
 557         if (evsel->cgrp)
 558                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 559
 560         if (csv_output || stat_config.interval)
 561                 return;
 562
 563         perf_stat__print_shadow_stats(output, evsel, avg, cpu,
 564                                       stat_config.aggr_mode);
 565 }
 566
 567 static void print_aggr(char *prefix)
 568 {
 569         FILE *output = stat_config.output;
 570         struct perf_evsel *counter;
 571         int cpu, s, s2, id, nr;
 572         double uval;
 573         u64 ena, run, val;
 574
 575         if (!(aggr_map || aggr_get_id))
 576                 return;
 577
 578         for (s = 0; s < aggr_map->nr; s++) {
 579                 id = aggr_map->map[s];
 580                 evlist__for_each(evsel_list, counter) {
 581                         val = ena = run = 0;
 582                         nr = 0;
 583                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 584                                 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
 585                                 if (s2 != id)
 586                                         continue;
 587                                 val += perf_counts(counter->counts, cpu, 0)->val;
 588                                 ena += perf_counts(counter->counts, cpu, 0)->ena;
 589                                 run += perf_counts(counter->counts, cpu, 0)->run;
 590                                 nr++;
 591                         }
 592                         if (prefix)
 593                                 fprintf(output, "%s", prefix);
 594
 595                         if (run == 0 || ena == 0) {
 596                                 aggr_printout(counter, id, nr);
 597
 598                                 fprintf(output, "%*s%s",
 599                                         csv_output ? 0 : 18,
 600                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 601                                         csv_sep);
 602
 603                                 fprintf(output, "%-*s%s",
 604                                         csv_output ? 0 : unit_width,
 605                                         counter->unit, csv_sep);
 606
 607                                 fprintf(output, "%*s",
 608                                         csv_output ? 0 : -25,
 609                                         perf_evsel__name(counter));
 610
 611                                 if (counter->cgrp)
 612                                         fprintf(output, "%s%s",
 613                                                 csv_sep, counter->cgrp->name);
 614
 615                                 print_running(run, ena);
 616                                 fputc('\n', output);
 617                                 continue;
 618                         }
 619                         uval = val * counter->scale;
 620
 621                         if (nsec_counter(counter))
 622                                 nsec_printout(id, nr, counter, uval);
 623                         else
 624                                 abs_printout(id, nr, counter, uval);
 625
 626                         if (!csv_output)
 627                                 print_noise(counter, 1.0);
 628
 629                         print_running(run, ena);
 630                         fputc('\n', output);
 631                 }
 632         }
 633 }
 634
 635 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
 636 {
 637         FILE *output = stat_config.output;
 638         int nthreads = thread_map__nr(counter->threads);
 639         int ncpus = cpu_map__nr(counter->cpus);
 640         int cpu, thread;
 641         double uval;
 642
 643         for (thread = 0; thread < nthreads; thread++) {
 644                 u64 ena = 0, run = 0, val = 0;
 645
 646                 for (cpu = 0; cpu < ncpus; cpu++) {
 647                         val += perf_counts(counter->counts, cpu, thread)->val;
 648                         ena += perf_counts(counter->counts, cpu, thread)->ena;
 649                         run += perf_counts(counter->counts, cpu, thread)->run;
 650                 }
 651
 652                 if (prefix)
 653                         fprintf(output, "%s", prefix);
 654
 655                 uval = val * counter->scale;
 656
 657                 if (nsec_counter(counter))
 658                         nsec_printout(thread, 0, counter, uval);
 659                 else
 660                         abs_printout(thread, 0, counter, uval);
 661
 662                 if (!csv_output)
 663                         print_noise(counter, 1.0);
 664
 665                 print_running(run, ena);
 666                 fputc('\n', output);
 667         }
 668 }
 669
 670 /*
 671  * Print out the results of a single counter:
 672  * aggregated counts in system-wide mode
 673  */
 674 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
 675 {
 676         FILE *output = stat_config.output;
 677         struct perf_stat_evsel *ps = counter->priv;
 678         double avg = avg_stats(&ps->res_stats[0]);
 679         int scaled = counter->counts->scaled;
 680         double uval;
 681         double avg_enabled, avg_running;
 682
 683         avg_enabled = avg_stats(&ps->res_stats[1]);
 684         avg_running = avg_stats(&ps->res_stats[2]);
 685
 686         if (prefix)
 687                 fprintf(output, "%s", prefix);
 688
 689         if (scaled == -1 || !counter->supported) {
 690                 fprintf(output, "%*s%s",
 691                         csv_output ? 0 : 18,
 692                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 693                         csv_sep);
 694                 fprintf(output, "%-*s%s",
 695                         csv_output ? 0 : unit_width,
 696                         counter->unit, csv_sep);
 697                 fprintf(output, "%*s",
 698                         csv_output ? 0 : -25,
 699                         perf_evsel__name(counter));
 700
 701                 if (counter->cgrp)
 702                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
 703
 704                 print_running(avg_running, avg_enabled);
 705                 fputc('\n', output);
 706                 return;
 707         }
 708
 709         uval = avg * counter->scale;
 710
 711         if (nsec_counter(counter))
 712                 nsec_printout(-1, 0, counter, uval);
 713         else
 714                 abs_printout(-1, 0, counter, uval);
 715
 716         print_noise(counter, avg);
 717
 718         print_running(avg_running, avg_enabled);
 719         fprintf(output, "\n");
 720 }
 721
 722 /*
 723  * Print out the results of a single counter:
 724  * does not use aggregated count in system-wide
 725  */
 726 static void print_counter(struct perf_evsel *counter, char *prefix)
 727 {
 728         FILE *output = stat_config.output;
 729         u64 ena, run, val;
 730         double uval;
 731         int cpu;
 732
 733         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 734                 val = perf_counts(counter->counts, cpu, 0)->val;
 735                 ena = perf_counts(counter->counts, cpu, 0)->ena;
 736                 run = perf_counts(counter->counts, cpu, 0)->run;
 737
 738                 if (prefix)
 739                         fprintf(output, "%s", prefix);
 740
 741                 if (run == 0 || ena == 0) {
 742                         fprintf(output, "CPU%*d%s%*s%s",
 743                                 csv_output ? 0 : -4,
 744                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
 745                                 csv_output ? 0 : 18,
 746                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 747                                 csv_sep);
 748
 749                                 fprintf(output, "%-*s%s",
 750                                         csv_output ? 0 : unit_width,
 751                                         counter->unit, csv_sep);
 752
 753                                 fprintf(output, "%*s",
 754                                         csv_output ? 0 : -25,
 755                                         perf_evsel__name(counter));
 756
 757                         if (counter->cgrp)
 758                                 fprintf(output, "%s%s",
 759                                         csv_sep, counter->cgrp->name);
 760
 761                         print_running(run, ena);
 762                         fputc('\n', output);
 763                         continue;
 764                 }
 765
 766                 uval = val * counter->scale;
 767
 768                 if (nsec_counter(counter))
 769                         nsec_printout(cpu, 0, counter, uval);
 770                 else
 771                         abs_printout(cpu, 0, counter, uval);
 772
 773                 if (!csv_output)
 774                         print_noise(counter, 1.0);
 775                 print_running(run, ena);
 776
 777                 fputc('\n', output);
 778         }
 779 }
 780
 781 static void print_interval(char *prefix, struct timespec *ts)
 782 {
 783         FILE *output = stat_config.output;
 784         static int num_print_interval;
 785
 786         sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
 787
 788         if (num_print_interval == 0 && !csv_output) {
 789                 switch (stat_config.aggr_mode) {
 790                 case AGGR_SOCKET:
 791                         fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
 792                         break;
 793                 case AGGR_CORE:
 794                         fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
 795                         break;
 796                 case AGGR_NONE:
 797                         fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
 798                         break;
 799                 case AGGR_THREAD:
 800                         fprintf(output, "#           time             comm-pid                  counts %*s events\n", unit_width, "unit");
 801                         break;
 802                 case AGGR_GLOBAL:
 803                 default:
 804                         fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
 805                 case AGGR_UNSET:
 806                         break;
 807                 }
 808         }
 809
 810         if (++num_print_interval == 25)
 811                 num_print_interval = 0;
 812 }
 813
 814 static void print_header(int argc, const char **argv)
 815 {
 816         FILE *output = stat_config.output;
 817         int i;
 818
 819         fflush(stdout);
 820
 821         if (!csv_output) {
 822                 fprintf(output, "\n");
 823                 fprintf(output, " Performance counter stats for ");
 824                 if (target.system_wide)
 825                         fprintf(output, "\'system wide");
 826                 else if (target.cpu_list)
 827                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
 828                 else if (!target__has_task(&target)) {
 829                         fprintf(output, "\'%s", argv[0]);
 830                         for (i = 1; i < argc; i++)
 831                                 fprintf(output, " %s", argv[i]);
 832                 } else if (target.pid)
 833                         fprintf(output, "process id \'%s", target.pid);
 834                 else
 835                         fprintf(output, "thread id \'%s", target.tid);
 836
 837                 fprintf(output, "\'");
 838                 if (run_count > 1)
 839                         fprintf(output, " (%d runs)", run_count);
 840                 fprintf(output, ":\n\n");
 841         }
 842 }
 843
 844 static void print_footer(void)
 845 {
 846         FILE *output = stat_config.output;
 847
 848         if (!null_run)
 849                 fprintf(output, "\n");
 850         fprintf(output, " %17.9f seconds time elapsed",
 851                         avg_stats(&walltime_nsecs_stats)/1e9);
 852         if (run_count > 1) {
 853                 fprintf(output, "                                        ");
 854                 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
 855                                 avg_stats(&walltime_nsecs_stats));
 856         }
 857         fprintf(output, "\n\n");
 858 }
 859
 860 static void print_counters(struct timespec *ts, int argc, const char **argv)
 861 {
 862         int interval = stat_config.interval;
 863         struct perf_evsel *counter;
 864         char buf[64], *prefix = NULL;
 865
 866         if (interval)
 867                 print_interval(prefix = buf, ts);
 868         else
 869                 print_header(argc, argv);
 870
 871         switch (stat_config.aggr_mode) {
 872         case AGGR_CORE:
 873         case AGGR_SOCKET:
 874                 print_aggr(prefix);
 875                 break;
 876         case AGGR_THREAD:
 877                 evlist__for_each(evsel_list, counter)
 878                         print_aggr_thread(counter, prefix);
 879                 break;
 880         case AGGR_GLOBAL:
 881                 evlist__for_each(evsel_list, counter)
 882                         print_counter_aggr(counter, prefix);
 883                 break;
 884         case AGGR_NONE:
 885                 evlist__for_each(evsel_list, counter)
 886                         print_counter(counter, prefix);
 887                 break;
 888         case AGGR_UNSET:
 889         default:
 890                 break;
 891         }
 892
 893         if (!interval && !csv_output)
 894                 print_footer();
 895
 896         fflush(stat_config.output);
 897 }
 898
 899 static volatile int signr = -1;
 900
 901 static void skip_signal(int signo)
 902 {
 903         if ((child_pid == -1) || stat_config.interval)
 904                 done = 1;
 905
 906         signr = signo;
 907         /*
 908          * render child_pid harmless
 909          * won't send SIGTERM to a random
 910          * process in case of race condition
 911          * and fast PID recycling
 912          */
 913         child_pid = -1;
 914 }
 915
 916 static void sig_atexit(void)
 917 {
 918         sigset_t set, oset;
 919
 920         /*
 921          * avoid race condition with SIGCHLD handler
 922          * in skip_signal() which is modifying child_pid
 923          * goal is to avoid send SIGTERM to a random
 924          * process
 925          */
 926         sigemptyset(&set);
 927         sigaddset(&set, SIGCHLD);
 928         sigprocmask(SIG_BLOCK, &set, &oset);
 929
 930         if (child_pid != -1)
 931                 kill(child_pid, SIGTERM);
 932
 933         sigprocmask(SIG_SETMASK, &oset, NULL);
 934
 935         if (signr == -1)
 936                 return;
 937
 938         signal(signr, SIG_DFL);
 939         kill(getpid(), signr);
 940 }
 941
 942 static int stat__set_big_num(const struct option *opt __maybe_unused,
 943                              const char *s __maybe_unused, int unset)
 944 {
 945         big_num_opt = unset ? 0 : 1;
 946         return 0;
 947 }
 948
 949 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
 950 {
 951         return cpu_map__get_socket(map, cpu, NULL);
 952 }
 953
 954 static int perf_stat__get_core(struct cpu_map *map, int cpu)
 955 {
 956         return cpu_map__get_core(map, cpu, NULL);
 957 }
 958
 959 static int cpu_map__get_max(struct cpu_map *map)
 960 {
 961         int i, max = -1;
 962
 963         for (i = 0; i < map->nr; i++) {
 964                 if (map->map[i] > max)
 965                         max = map->map[i];
 966         }
 967
 968         return max;
 969 }
 970
 971 static struct cpu_map *cpus_aggr_map;
 972
 973 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
 974 {
 975         int cpu;
 976
 977         if (idx >= map->nr)
 978                 return -1;
 979
 980         cpu = map->map[idx];
 981
 982         if (cpus_aggr_map->map[cpu] == -1)
 983                 cpus_aggr_map->map[cpu] = get_id(map, idx);
 984
 985         return cpus_aggr_map->map[cpu];
 986 }
 987
 988 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
 989 {
 990         return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
 991 }
 992
 993 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
 994 {
 995         return perf_stat__get_aggr(perf_stat__get_core, map, idx);
 996 }
 997
 998 static int perf_stat_init_aggr_mode(void)
 999 {
1000         int nr;
1001
1002         switch (stat_config.aggr_mode) {
1003         case AGGR_SOCKET:
1004                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1005                         perror("cannot build socket map");
1006                         return -1;
1007                 }
1008                 aggr_get_id = perf_stat__get_socket_cached;
1009                 break;
1010         case AGGR_CORE:
1011                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1012                         perror("cannot build core map");
1013                         return -1;
1014                 }
1015                 aggr_get_id = perf_stat__get_core_cached;
1016                 break;
1017         case AGGR_NONE:
1018         case AGGR_GLOBAL:
1019         case AGGR_THREAD:
1020         case AGGR_UNSET:
1021         default:
1022                 break;
1023         }
1024
1025         /*
1026          * The evsel_list->cpus is the base we operate on,
1027          * taking the highest cpu number to be the size of
1028          * the aggregation translate cpumap.
1029          */
1030         nr = cpu_map__get_max(evsel_list->cpus);
1031         cpus_aggr_map = cpu_map__empty_new(nr + 1);
1032         return cpus_aggr_map ? 0 : -ENOMEM;
1033 }
1034
1035 /*
1036  * Add default attributes, if there were no attributes specified or
1037  * if -d/--detailed, -d -d or -d -d -d is used:
1038  */
1039 static int add_default_attributes(void)
1040 {
1041         struct perf_event_attr default_attrs[] = {
1042
1043   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1044   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1045   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1046   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1047
1048   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1049   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1050   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1051   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1052   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1053   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1054
1055 };
1056
1057 /*
1058  * Detailed stats (-d), covering the L1 and last level data caches:
1059  */
1060         struct perf_event_attr detailed_attrs[] = {
1061
1062   { .type = PERF_TYPE_HW_CACHE,
1063     .config =
1064          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1065         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1066         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1067
1068   { .type = PERF_TYPE_HW_CACHE,
1069     .config =
1070          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1071         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1072         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1073
1074   { .type = PERF_TYPE_HW_CACHE,
1075     .config =
1076          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1077         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1078         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1079
1080   { .type = PERF_TYPE_HW_CACHE,
1081     .config =
1082          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1083         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1084         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1085 };
1086
1087 /*
1088  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1089  */
1090         struct perf_event_attr very_detailed_attrs[] = {
1091
1092   { .type = PERF_TYPE_HW_CACHE,
1093     .config =
1094          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1095         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1096         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1097
1098   { .type = PERF_TYPE_HW_CACHE,
1099     .config =
1100          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1101         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1102         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1103
1104   { .type = PERF_TYPE_HW_CACHE,
1105     .config =
1106          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1107         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1108         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1109
1110   { .type = PERF_TYPE_HW_CACHE,
1111     .config =
1112          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1113         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1114         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1115
1116   { .type = PERF_TYPE_HW_CACHE,
1117     .config =
1118          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1119         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1120         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1121
1122   { .type = PERF_TYPE_HW_CACHE,
1123     .config =
1124          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1125         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1126         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1127
1128 };
1129
1130 /*
1131  * Very, very detailed stats (-d -d -d), adding prefetch events:
1132  */
1133         struct perf_event_attr very_very_detailed_attrs[] = {
1134
1135   { .type = PERF_TYPE_HW_CACHE,
1136     .config =
1137          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1138         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1139         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1140
1141   { .type = PERF_TYPE_HW_CACHE,
1142     .config =
1143          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1144         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1145         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1146 };
1147
1148         /* Set attrs if no event is selected and !null_run: */
1149         if (null_run)
1150                 return 0;
1151
1152         if (transaction_run) {
1153                 int err;
1154                 if (pmu_have_event("cpu", "cycles-ct") &&
1155                     pmu_have_event("cpu", "el-start"))
1156                         err = parse_events(evsel_list, transaction_attrs, NULL);
1157                 else
1158                         err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1159                 if (err) {
1160                         fprintf(stderr, "Cannot set up transaction events\n");
1161                         return -1;
1162                 }
1163                 return 0;
1164         }
1165
1166         if (!evsel_list->nr_entries) {
1167                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1168                         return -1;
1169         }
1170
1171         /* Detailed events get appended to the event list: */
1172
1173         if (detailed_run <  1)
1174                 return 0;
1175
1176         /* Append detailed run extra attributes: */
1177         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1178                 return -1;
1179
1180         if (detailed_run < 2)
1181                 return 0;
1182
1183         /* Append very detailed run extra attributes: */
1184         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1185                 return -1;
1186
1187         if (detailed_run < 3)
1188                 return 0;
1189
1190         /* Append very, very detailed run extra attributes: */
1191         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1192 }
1193
1194 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1195 {
1196         bool append_file = false;
1197         int output_fd = 0;
1198         const char *output_name = NULL;
1199         const struct option options[] = {
1200         OPT_BOOLEAN('T', "transaction", &transaction_run,
1201                     "hardware transaction statistics"),
1202         OPT_CALLBACK('e', "event", &evsel_list, "event",
1203                      "event selector. use 'perf list' to list available events",
1204                      parse_events_option),
1205         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1206                      "event filter", parse_filter),
1207         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1208                     "child tasks do not inherit counters"),
1209         OPT_STRING('p', "pid", &target.pid, "pid",
1210                    "stat events on existing process id"),
1211         OPT_STRING('t', "tid", &target.tid, "tid",
1212                    "stat events on existing thread id"),
1213         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1214                     "system-wide collection from all CPUs"),
1215         OPT_BOOLEAN('g', "group", &group,
1216                     "put the counters into a counter group"),
1217         OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1218         OPT_INCR('v', "verbose", &verbose,
1219                     "be more verbose (show counter open errors, etc)"),
1220         OPT_INTEGER('r', "repeat", &run_count,
1221                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1222         OPT_BOOLEAN('n', "null", &null_run,
1223                     "null run - dont start any counters"),
1224         OPT_INCR('d', "detailed", &detailed_run,
1225                     "detailed run - start a lot of events"),
1226         OPT_BOOLEAN('S', "sync", &sync_run,
1227                     "call sync() before starting a run"),
1228         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1229                            "print large numbers with thousands\' separators",
1230                            stat__set_big_num),
1231         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1232                     "list of cpus to monitor in system-wide"),
1233         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1234                     "disable CPU count aggregation", AGGR_NONE),
1235         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1236                    "print counts with custom separator"),
1237         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1238                      "monitor event in cgroup name only", parse_cgroups),
1239         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1240         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1241         OPT_INTEGER(0, "log-fd", &output_fd,
1242                     "log output to fd, instead of stderr"),
1243         OPT_STRING(0, "pre", &pre_cmd, "command",
1244                         "command to run prior to the measured command"),
1245         OPT_STRING(0, "post", &post_cmd, "command",
1246                         "command to run after to the measured command"),
1247         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1248                     "print counts at regular interval in ms (>= 10)"),
1249         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1250                      "aggregate counts per processor socket", AGGR_SOCKET),
1251         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1252                      "aggregate counts per physical processor core", AGGR_CORE),
1253         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1254                      "aggregate counts per thread", AGGR_THREAD),
1255         OPT_UINTEGER('D', "delay", &initial_delay,
1256                      "ms to wait before starting measurement after program start"),
1257         OPT_END()
1258         };
1259         const char * const stat_usage[] = {
1260                 "perf stat [<options>] [<command>]",
1261                 NULL
1262         };
1263         int status = -EINVAL, run_idx;
1264         const char *mode;
1265         FILE *output = stderr;
1266         unsigned int interval;
1267
1268         setlocale(LC_ALL, "");
1269
1270         evsel_list = perf_evlist__new();
1271         if (evsel_list == NULL)
1272                 return -ENOMEM;
1273
1274         argc = parse_options(argc, argv, options, stat_usage,
1275                 PARSE_OPT_STOP_AT_NON_OPTION);
1276
1277         interval = stat_config.interval;
1278
1279         if (output_name && strcmp(output_name, "-"))
1280                 output = NULL;
1281
1282         if (output_name && output_fd) {
1283                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1284                 parse_options_usage(stat_usage, options, "o", 1);
1285                 parse_options_usage(NULL, options, "log-fd", 0);
1286                 goto out;
1287         }
1288
1289         if (output_fd < 0) {
1290                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1291                 parse_options_usage(stat_usage, options, "log-fd", 0);
1292                 goto out;
1293         }
1294
1295         if (!output) {
1296                 struct timespec tm;
1297                 mode = append_file ? "a" : "w";
1298
1299                 output = fopen(output_name, mode);
1300                 if (!output) {
1301                         perror("failed to create output file");
1302                         return -1;
1303                 }
1304                 clock_gettime(CLOCK_REALTIME, &tm);
1305                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1306         } else if (output_fd > 0) {
1307                 mode = append_file ? "a" : "w";
1308                 output = fdopen(output_fd, mode);
1309                 if (!output) {
1310                         perror("Failed opening logfd");
1311                         return -errno;
1312                 }
1313         }
1314
1315         stat_config.output = output;
1316
1317         if (csv_sep) {
1318                 csv_output = true;
1319                 if (!strcmp(csv_sep, "\\t"))
1320                         csv_sep = "\t";
1321         } else
1322                 csv_sep = DEFAULT_SEPARATOR;
1323
1324         /*
1325          * let the spreadsheet do the pretty-printing
1326          */
1327         if (csv_output) {
1328                 /* User explicitly passed -B? */
1329                 if (big_num_opt == 1) {
1330                         fprintf(stderr, "-B option not supported with -x\n");
1331                         parse_options_usage(stat_usage, options, "B", 1);
1332                         parse_options_usage(NULL, options, "x", 1);
1333                         goto out;
1334                 } else /* Nope, so disable big number formatting */
1335                         big_num = false;
1336         } else if (big_num_opt == 0) /* User passed --no-big-num */
1337                 big_num = false;
1338
1339         if (!argc && target__none(&target))
1340                 usage_with_options(stat_usage, options);
1341
1342         if (run_count < 0) {
1343                 pr_err("Run count must be a positive number\n");
1344                 parse_options_usage(stat_usage, options, "r", 1);
1345                 goto out;
1346         } else if (run_count == 0) {
1347                 forever = true;
1348                 run_count = 1;
1349         }
1350
1351         if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1352                 fprintf(stderr, "The --per-thread option is only available "
1353                         "when monitoring via -p -t options.\n");
1354                 parse_options_usage(NULL, options, "p", 1);
1355                 parse_options_usage(NULL, options, "t", 1);
1356                 goto out;
1357         }
1358
1359         /*
1360          * no_aggr, cgroup are for system-wide only
1361          * --per-thread is aggregated per thread, we dont mix it with cpu mode
1362          */
1363         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1364               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1365             !target__has_cpu(&target)) {
1366                 fprintf(stderr, "both cgroup and no-aggregation "
1367                         "modes only available in system-wide mode\n");
1368
1369                 parse_options_usage(stat_usage, options, "G", 1);
1370                 parse_options_usage(NULL, options, "A", 1);
1371                 parse_options_usage(NULL, options, "a", 1);
1372                 goto out;
1373         }
1374
1375         if (add_default_attributes())
1376                 goto out;
1377
1378         target__validate(&target);
1379
1380         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1381                 if (target__has_task(&target)) {
1382                         pr_err("Problems finding threads of monitor\n");
1383                         parse_options_usage(stat_usage, options, "p", 1);
1384                         parse_options_usage(NULL, options, "t", 1);
1385                 } else if (target__has_cpu(&target)) {
1386                         perror("failed to parse CPUs map");
1387                         parse_options_usage(stat_usage, options, "C", 1);
1388                         parse_options_usage(NULL, options, "a", 1);
1389                 }
1390                 goto out;
1391         }
1392
1393         /*
1394          * Initialize thread_map with comm names,
1395          * so we could print it out on output.
1396          */
1397         if (stat_config.aggr_mode == AGGR_THREAD)
1398                 thread_map__read_comms(evsel_list->threads);
1399
1400         if (interval && interval < 100) {
1401                 if (interval < 10) {
1402                         pr_err("print interval must be >= 10ms\n");
1403                         parse_options_usage(stat_usage, options, "I", 1);
1404                         goto out;
1405                 } else
1406                         pr_warning("print interval < 100ms. "
1407                                    "The overhead percentage could be high in some cases. "
1408                                    "Please proceed with caution.\n");
1409         }
1410
1411         if (perf_evlist__alloc_stats(evsel_list, interval))
1412                 goto out;
1413
1414         if (perf_stat_init_aggr_mode())
1415                 goto out;
1416
1417         /*
1418          * We dont want to block the signals - that would cause
1419          * child tasks to inherit that and Ctrl-C would not work.
1420          * What we want is for Ctrl-C to work in the exec()-ed
1421          * task, but being ignored by perf stat itself:
1422          */
1423         atexit(sig_atexit);
1424         if (!forever)
1425                 signal(SIGINT,  skip_signal);
1426         signal(SIGCHLD, skip_signal);
1427         signal(SIGALRM, skip_signal);
1428         signal(SIGABRT, skip_signal);
1429
1430         status = 0;
1431         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1432                 if (run_count != 1 && verbose)
1433                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1434                                 run_idx + 1);
1435
1436                 status = run_perf_stat(argc, argv);
1437                 if (forever && status != -1) {
1438                         print_counters(NULL, argc, argv);
1439                         perf_stat__reset_stats();
1440                 }
1441         }
1442
1443         if (!forever && status != -1 && !interval)
1444                 print_counters(NULL, argc, argv);
1445
1446         perf_evlist__free_stats(evsel_list);
1447 out:
1448         perf_evlist__delete(evsel_list);
1449         return status;
1450 }