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
  62 #include <stdlib.h>
  63 #include <sys/prctl.h>
  64 #include <locale.h>
  65
  66 #define DEFAULT_SEPARATOR       " "
  67 #define CNTR_NOT_SUPPORTED      "<not supported>"
  68 #define CNTR_NOT_COUNTED        "<not counted>"
  69
  70 static void print_stat(int argc, const char **argv);
  71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
  72 static void print_counter(struct perf_evsel *counter, char *prefix);
  73 static void print_aggr(char *prefix);
  74
  75 /* Default events used for perf stat -T */
  76 static const char *transaction_attrs = {
  77         "task-clock,"
  78         "{"
  79         "instructions,"
  80         "cycles,"
  81         "cpu/cycles-t/,"
  82         "cpu/tx-start/,"
  83         "cpu/el-start/,"
  84         "cpu/cycles-ct/"
  85         "}"
  86 };
  87
  88 /* More limited version when the CPU does not have all events. */
  89 static const char * transaction_limited_attrs = {
  90         "task-clock,"
  91         "{"
  92         "instructions,"
  93         "cycles,"
  94         "cpu/cycles-t/,"
  95         "cpu/tx-start/"
  96         "}"
  97 };
  98
  99 static struct perf_evlist       *evsel_list;
 100
 101 static struct target target = {
 102         .uid    = UINT_MAX,
 103 };
 104
 105 static int                      run_count                       =  1;
 106 static bool                     no_inherit                      = false;
 107 static bool                     scale                           =  true;
 108 static enum aggr_mode           aggr_mode                       = AGGR_GLOBAL;
 109 static volatile pid_t           child_pid                       = -1;
 110 static bool                     null_run                        =  false;
 111 static int                      detailed_run                    =  0;
 112 static bool                     transaction_run;
 113 static bool                     big_num                         =  true;
 114 static int                      big_num_opt                     =  -1;
 115 static const char               *csv_sep                        = NULL;
 116 static bool                     csv_output                      = false;
 117 static bool                     group                           = false;
 118 static FILE                     *output                         = NULL;
 119 static const char               *pre_cmd                        = NULL;
 120 static const char               *post_cmd                       = NULL;
 121 static bool                     sync_run                        = false;
 122 static unsigned int             interval                        = 0;
 123 static unsigned int             initial_delay                   = 0;
 124 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 125 static bool                     forever                         = false;
 126 static struct timespec          ref_time;
 127 static struct cpu_map           *aggr_map;
 128 static int                      (*aggr_get_id)(struct cpu_map *m, int cpu);
 129
 130 static volatile int done = 0;
 131
 132 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 133                                  struct timespec *b)
 134 {
 135         r->tv_sec = a->tv_sec - b->tv_sec;
 136         if (a->tv_nsec < b->tv_nsec) {
 137                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 138                 r->tv_sec--;
 139         } else {
 140                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 141         }
 142 }
 143
 144 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 145 {
 146         int i;
 147         struct perf_stat *ps = evsel->priv;
 148
 149         for (i = 0; i < 3; i++)
 150                 init_stats(&ps->res_stats[i]);
 151
 152         perf_stat_evsel_id_init(evsel);
 153 }
 154
 155 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 156 {
 157         evsel->priv = zalloc(sizeof(struct perf_stat));
 158         if (evsel->priv == NULL)
 159                 return -ENOMEM;
 160         perf_evsel__reset_stat_priv(evsel);
 161         return 0;
 162 }
 163
 164 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 165 {
 166         zfree(&evsel->priv);
 167 }
 168
 169 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
 170 {
 171         struct perf_counts *counts;
 172
 173         counts = perf_counts__new(perf_evsel__nr_cpus(evsel));
 174         if (counts)
 175                 evsel->prev_raw_counts = counts;
 176
 177         return counts ? 0 : -ENOMEM;
 178 }
 179
 180 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
 181 {
 182         perf_counts__delete(evsel->prev_raw_counts);
 183         evsel->prev_raw_counts = NULL;
 184 }
 185
 186 static void perf_evlist__free_stats(struct perf_evlist *evlist)
 187 {
 188         struct perf_evsel *evsel;
 189
 190         evlist__for_each(evlist, evsel) {
 191                 perf_evsel__free_stat_priv(evsel);
 192                 perf_evsel__free_counts(evsel);
 193                 perf_evsel__free_prev_raw_counts(evsel);
 194         }
 195 }
 196
 197 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
 198 {
 199         struct perf_evsel *evsel;
 200
 201         evlist__for_each(evlist, evsel) {
 202                 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 203                     perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
 204                     (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
 205                         goto out_free;
 206         }
 207
 208         return 0;
 209
 210 out_free:
 211         perf_evlist__free_stats(evlist);
 212         return -1;
 213 }
 214
 215 static void perf_stat__reset_stats(struct perf_evlist *evlist)
 216 {
 217         struct perf_evsel *evsel;
 218
 219         evlist__for_each(evlist, evsel) {
 220                 perf_evsel__reset_stat_priv(evsel);
 221                 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
 222         }
 223
 224         perf_stat__reset_shadow_stats();
 225 }
 226
 227 static int create_perf_stat_counter(struct perf_evsel *evsel)
 228 {
 229         struct perf_event_attr *attr = &evsel->attr;
 230
 231         if (scale)
 232                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 233                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 234
 235         attr->inherit = !no_inherit;
 236
 237         if (target__has_cpu(&target))
 238                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 239
 240         if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
 241                 attr->disabled = 1;
 242                 if (!initial_delay)
 243                         attr->enable_on_exec = 1;
 244         }
 245
 246         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 247 }
 248
 249 /*
 250  * Does the counter have nsecs as a unit?
 251  */
 252 static inline int nsec_counter(struct perf_evsel *evsel)
 253 {
 254         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 255             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 256                 return 1;
 257
 258         return 0;
 259 }
 260
 261 static void zero_per_pkg(struct perf_evsel *counter)
 262 {
 263         if (counter->per_pkg_mask)
 264                 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
 265 }
 266
 267 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
 268 {
 269         unsigned long *mask = counter->per_pkg_mask;
 270         struct cpu_map *cpus = perf_evsel__cpus(counter);
 271         int s;
 272
 273         *skip = false;
 274
 275         if (!counter->per_pkg)
 276                 return 0;
 277
 278         if (cpu_map__empty(cpus))
 279                 return 0;
 280
 281         if (!mask) {
 282                 mask = zalloc(MAX_NR_CPUS);
 283                 if (!mask)
 284                         return -ENOMEM;
 285
 286                 counter->per_pkg_mask = mask;
 287         }
 288
 289         s = cpu_map__get_socket(cpus, cpu);
 290         if (s < 0)
 291                 return -1;
 292
 293         *skip = test_and_set_bit(s, mask) == 1;
 294         return 0;
 295 }
 296
 297 static int read_cb(struct perf_evsel *evsel, int cpu, int thread __maybe_unused,
 298                    struct perf_counts_values *count)
 299 {
 300         struct perf_counts_values *aggr = &evsel->counts->aggr;
 301         static struct perf_counts_values zero;
 302         bool skip = false;
 303
 304         if (check_per_pkg(evsel, cpu, &skip)) {
 305                 pr_err("failed to read per-pkg counter\n");
 306                 return -1;
 307         }
 308
 309         if (skip)
 310                 count = &zero;
 311
 312         switch (aggr_mode) {
 313         case AGGR_CORE:
 314         case AGGR_SOCKET:
 315         case AGGR_NONE:
 316                 if (!evsel->snapshot)
 317                         perf_evsel__compute_deltas(evsel, cpu, count);
 318                 perf_counts_values__scale(count, scale, NULL);
 319                 *perf_counts(evsel->counts, cpu) = *count;
 320                 if (aggr_mode == AGGR_NONE)
 321                         perf_stat__update_shadow_stats(evsel, count->values, cpu);
 322                 break;
 323         case AGGR_GLOBAL:
 324                 aggr->val += count->val;
 325                 if (scale) {
 326                         aggr->ena += count->ena;
 327                         aggr->run += count->run;
 328                 }
 329         default:
 330                 break;
 331         }
 332
 333         return 0;
 334 }
 335
 336 static int read_counter(struct perf_evsel *counter);
 337
 338 /*
 339  * Read out the results of a single counter:
 340  * aggregate counts across CPUs in system-wide mode
 341  */
 342 static int read_counter_aggr(struct perf_evsel *counter)
 343 {
 344         struct perf_counts_values *aggr = &counter->counts->aggr;
 345         struct perf_stat *ps = counter->priv;
 346         u64 *count = counter->counts->aggr.values;
 347         int i;
 348
 349         aggr->val = aggr->ena = aggr->run = 0;
 350
 351         if (read_counter(counter))
 352                 return -1;
 353
 354         if (!counter->snapshot)
 355                 perf_evsel__compute_deltas(counter, -1, aggr);
 356         perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
 357
 358         for (i = 0; i < 3; i++)
 359                 update_stats(&ps->res_stats[i], count[i]);
 360
 361         if (verbose) {
 362                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 363                         perf_evsel__name(counter), count[0], count[1], count[2]);
 364         }
 365
 366         /*
 367          * Save the full runtime - to allow normalization during printout:
 368          */
 369         perf_stat__update_shadow_stats(counter, count, 0);
 370
 371         return 0;
 372 }
 373
 374 /*
 375  * Read out the results of a single counter:
 376  * do not aggregate counts across CPUs in system-wide mode
 377  */
 378 static int read_counter(struct perf_evsel *counter)
 379 {
 380         int nthreads = thread_map__nr(evsel_list->threads);
 381         int ncpus = perf_evsel__nr_cpus(counter);
 382         int cpu, thread;
 383
 384         if (!counter->supported)
 385                 return -ENOENT;
 386
 387         if (counter->system_wide)
 388                 nthreads = 1;
 389
 390         if (counter->per_pkg)
 391                 zero_per_pkg(counter);
 392
 393         for (thread = 0; thread < nthreads; thread++) {
 394                 for (cpu = 0; cpu < ncpus; cpu++) {
 395                         if (perf_evsel__read_cb(counter, cpu, thread, read_cb))
 396                                 return -1;
 397                 }
 398         }
 399
 400         return 0;
 401 }
 402
 403 static void print_interval(void)
 404 {
 405         static int num_print_interval;
 406         struct perf_evsel *counter;
 407         struct perf_stat *ps;
 408         struct timespec ts, rs;
 409         char prefix[64];
 410
 411         if (aggr_mode == AGGR_GLOBAL) {
 412                 evlist__for_each(evsel_list, counter) {
 413                         ps = counter->priv;
 414                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
 415                         read_counter_aggr(counter);
 416                 }
 417         } else  {
 418                 evlist__for_each(evsel_list, counter) {
 419                         ps = counter->priv;
 420                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
 421                         read_counter(counter);
 422                 }
 423         }
 424
 425         clock_gettime(CLOCK_MONOTONIC, &ts);
 426         diff_timespec(&rs, &ts, &ref_time);
 427         sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
 428
 429         if (num_print_interval == 0 && !csv_output) {
 430                 switch (aggr_mode) {
 431                 case AGGR_SOCKET:
 432                         fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
 433                         break;
 434                 case AGGR_CORE:
 435                         fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
 436                         break;
 437                 case AGGR_NONE:
 438                         fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
 439                         break;
 440                 case AGGR_GLOBAL:
 441                 default:
 442                         fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
 443                 }
 444         }
 445
 446         if (++num_print_interval == 25)
 447                 num_print_interval = 0;
 448
 449         switch (aggr_mode) {
 450         case AGGR_CORE:
 451         case AGGR_SOCKET:
 452                 print_aggr(prefix);
 453                 break;
 454         case AGGR_NONE:
 455                 evlist__for_each(evsel_list, counter)
 456                         print_counter(counter, prefix);
 457                 break;
 458         case AGGR_GLOBAL:
 459         default:
 460                 evlist__for_each(evsel_list, counter)
 461                         print_counter_aggr(counter, prefix);
 462         }
 463
 464         fflush(output);
 465 }
 466
 467 static void handle_initial_delay(void)
 468 {
 469         struct perf_evsel *counter;
 470
 471         if (initial_delay) {
 472                 const int ncpus = cpu_map__nr(evsel_list->cpus),
 473                         nthreads = thread_map__nr(evsel_list->threads);
 474
 475                 usleep(initial_delay * 1000);
 476                 evlist__for_each(evsel_list, counter)
 477                         perf_evsel__enable(counter, ncpus, nthreads);
 478         }
 479 }
 480
 481 static volatile int workload_exec_errno;
 482
 483 /*
 484  * perf_evlist__prepare_workload will send a SIGUSR1
 485  * if the fork fails, since we asked by setting its
 486  * want_signal to true.
 487  */
 488 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 489                                         void *ucontext __maybe_unused)
 490 {
 491         workload_exec_errno = info->si_value.sival_int;
 492 }
 493
 494 static int __run_perf_stat(int argc, const char **argv)
 495 {
 496         char msg[512];
 497         unsigned long long t0, t1;
 498         struct perf_evsel *counter;
 499         struct timespec ts;
 500         size_t l;
 501         int status = 0;
 502         const bool forks = (argc > 0);
 503
 504         if (interval) {
 505                 ts.tv_sec  = interval / 1000;
 506                 ts.tv_nsec = (interval % 1000) * 1000000;
 507         } else {
 508                 ts.tv_sec  = 1;
 509                 ts.tv_nsec = 0;
 510         }
 511
 512         if (forks) {
 513                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
 514                                                   workload_exec_failed_signal) < 0) {
 515                         perror("failed to prepare workload");
 516                         return -1;
 517                 }
 518                 child_pid = evsel_list->workload.pid;
 519         }
 520
 521         if (group)
 522                 perf_evlist__set_leader(evsel_list);
 523
 524         evlist__for_each(evsel_list, counter) {
 525                 if (create_perf_stat_counter(counter) < 0) {
 526                         /*
 527                          * PPC returns ENXIO for HW counters until 2.6.37
 528                          * (behavior changed with commit b0a873e).
 529                          */
 530                         if (errno == EINVAL || errno == ENOSYS ||
 531                             errno == ENOENT || errno == EOPNOTSUPP ||
 532                             errno == ENXIO) {
 533                                 if (verbose)
 534                                         ui__warning("%s event is not supported by the kernel.\n",
 535                                                     perf_evsel__name(counter));
 536                                 counter->supported = false;
 537
 538                                 if ((counter->leader != counter) ||
 539                                     !(counter->leader->nr_members > 1))
 540                                         continue;
 541                         }
 542
 543                         perf_evsel__open_strerror(counter, &target,
 544                                                   errno, msg, sizeof(msg));
 545                         ui__error("%s\n", msg);
 546
 547                         if (child_pid != -1)
 548                                 kill(child_pid, SIGTERM);
 549
 550                         return -1;
 551                 }
 552                 counter->supported = true;
 553
 554                 l = strlen(counter->unit);
 555                 if (l > unit_width)
 556                         unit_width = l;
 557         }
 558
 559         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 560                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 561                         counter->filter, perf_evsel__name(counter), errno,
 562                         strerror_r(errno, msg, sizeof(msg)));
 563                 return -1;
 564         }
 565
 566         /*
 567          * Enable counters and exec the command:
 568          */
 569         t0 = rdclock();
 570         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 571
 572         if (forks) {
 573                 perf_evlist__start_workload(evsel_list);
 574                 handle_initial_delay();
 575
 576                 if (interval) {
 577                         while (!waitpid(child_pid, &status, WNOHANG)) {
 578                                 nanosleep(&ts, NULL);
 579                                 print_interval();
 580                         }
 581                 }
 582                 wait(&status);
 583
 584                 if (workload_exec_errno) {
 585                         const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 586                         pr_err("Workload failed: %s\n", emsg);
 587                         return -1;
 588                 }
 589
 590                 if (WIFSIGNALED(status))
 591                         psignal(WTERMSIG(status), argv[0]);
 592         } else {
 593                 handle_initial_delay();
 594                 while (!done) {
 595                         nanosleep(&ts, NULL);
 596                         if (interval)
 597                                 print_interval();
 598                 }
 599         }
 600
 601         t1 = rdclock();
 602
 603         update_stats(&walltime_nsecs_stats, t1 - t0);
 604
 605         if (aggr_mode == AGGR_GLOBAL) {
 606                 evlist__for_each(evsel_list, counter) {
 607                         read_counter_aggr(counter);
 608                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 609                                              thread_map__nr(evsel_list->threads));
 610                 }
 611         } else {
 612                 evlist__for_each(evsel_list, counter) {
 613                         read_counter(counter);
 614                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
 615                 }
 616         }
 617
 618         return WEXITSTATUS(status);
 619 }
 620
 621 static int run_perf_stat(int argc, const char **argv)
 622 {
 623         int ret;
 624
 625         if (pre_cmd) {
 626                 ret = system(pre_cmd);
 627                 if (ret)
 628                         return ret;
 629         }
 630
 631         if (sync_run)
 632                 sync();
 633
 634         ret = __run_perf_stat(argc, argv);
 635         if (ret)
 636                 return ret;
 637
 638         if (post_cmd) {
 639                 ret = system(post_cmd);
 640                 if (ret)
 641                         return ret;
 642         }
 643
 644         return ret;
 645 }
 646
 647 static void print_running(u64 run, u64 ena)
 648 {
 649         if (csv_output) {
 650                 fprintf(output, "%s%" PRIu64 "%s%.2f",
 651                                         csv_sep,
 652                                         run,
 653                                         csv_sep,
 654                                         ena ? 100.0 * run / ena : 100.0);
 655         } else if (run != ena) {
 656                 fprintf(output, "  (%.2f%%)", 100.0 * run / ena);
 657         }
 658 }
 659
 660 static void print_noise_pct(double total, double avg)
 661 {
 662         double pct = rel_stddev_stats(total, avg);
 663
 664         if (csv_output)
 665                 fprintf(output, "%s%.2f%%", csv_sep, pct);
 666         else if (pct)
 667                 fprintf(output, "  ( +-%6.2f%% )", pct);
 668 }
 669
 670 static void print_noise(struct perf_evsel *evsel, double avg)
 671 {
 672         struct perf_stat *ps;
 673
 674         if (run_count == 1)
 675                 return;
 676
 677         ps = evsel->priv;
 678         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 679 }
 680
 681 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 682 {
 683         switch (aggr_mode) {
 684         case AGGR_CORE:
 685                 fprintf(output, "S%d-C%*d%s%*d%s",
 686                         cpu_map__id_to_socket(id),
 687                         csv_output ? 0 : -8,
 688                         cpu_map__id_to_cpu(id),
 689                         csv_sep,
 690                         csv_output ? 0 : 4,
 691                         nr,
 692                         csv_sep);
 693                 break;
 694         case AGGR_SOCKET:
 695                 fprintf(output, "S%*d%s%*d%s",
 696                         csv_output ? 0 : -5,
 697                         id,
 698                         csv_sep,
 699                         csv_output ? 0 : 4,
 700                         nr,
 701                         csv_sep);
 702                         break;
 703         case AGGR_NONE:
 704                 fprintf(output, "CPU%*d%s",
 705                         csv_output ? 0 : -4,
 706                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 707                 break;
 708         case AGGR_GLOBAL:
 709         default:
 710                 break;
 711         }
 712 }
 713
 714 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 715 {
 716         double msecs = avg / 1e6;
 717         const char *fmt_v, *fmt_n;
 718         char name[25];
 719
 720         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
 721         fmt_n = csv_output ? "%s" : "%-25s";
 722
 723         aggr_printout(evsel, id, nr);
 724
 725         scnprintf(name, sizeof(name), "%s%s",
 726                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
 727
 728         fprintf(output, fmt_v, msecs, csv_sep);
 729
 730         if (csv_output)
 731                 fprintf(output, "%s%s", evsel->unit, csv_sep);
 732         else
 733                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
 734
 735         fprintf(output, fmt_n, name);
 736
 737         if (evsel->cgrp)
 738                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 739
 740         if (csv_output || interval)
 741                 return;
 742
 743         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 744                 fprintf(output, " # %8.3f CPUs utilized          ",
 745                         avg / avg_stats(&walltime_nsecs_stats));
 746         else
 747                 fprintf(output, "                                   ");
 748 }
 749
 750 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 751 {
 752         double sc =  evsel->scale;
 753         const char *fmt;
 754         int cpu = cpu_map__id_to_cpu(id);
 755
 756         if (csv_output) {
 757                 fmt = sc != 1.0 ?  "%.2f%s" : "%.0f%s";
 758         } else {
 759                 if (big_num)
 760                         fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
 761                 else
 762                         fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
 763         }
 764
 765         aggr_printout(evsel, id, nr);
 766
 767         if (aggr_mode == AGGR_GLOBAL)
 768                 cpu = 0;
 769
 770         fprintf(output, fmt, avg, csv_sep);
 771
 772         if (evsel->unit)
 773                 fprintf(output, "%-*s%s",
 774                         csv_output ? 0 : unit_width,
 775                         evsel->unit, csv_sep);
 776
 777         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
 778
 779         if (evsel->cgrp)
 780                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 781
 782         if (csv_output || interval)
 783                 return;
 784
 785         perf_stat__print_shadow_stats(output, evsel, avg, cpu, aggr_mode);
 786 }
 787
 788 static void print_aggr(char *prefix)
 789 {
 790         struct perf_evsel *counter;
 791         int cpu, cpu2, s, s2, id, nr;
 792         double uval;
 793         u64 ena, run, val;
 794
 795         if (!(aggr_map || aggr_get_id))
 796                 return;
 797
 798         for (s = 0; s < aggr_map->nr; s++) {
 799                 id = aggr_map->map[s];
 800                 evlist__for_each(evsel_list, counter) {
 801                         val = ena = run = 0;
 802                         nr = 0;
 803                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 804                                 cpu2 = perf_evsel__cpus(counter)->map[cpu];
 805                                 s2 = aggr_get_id(evsel_list->cpus, cpu2);
 806                                 if (s2 != id)
 807                                         continue;
 808                                 val += perf_counts(counter->counts, cpu)->val;
 809                                 ena += perf_counts(counter->counts, cpu)->ena;
 810                                 run += perf_counts(counter->counts, cpu)->run;
 811                                 nr++;
 812                         }
 813                         if (prefix)
 814                                 fprintf(output, "%s", prefix);
 815
 816                         if (run == 0 || ena == 0) {
 817                                 aggr_printout(counter, id, nr);
 818
 819                                 fprintf(output, "%*s%s",
 820                                         csv_output ? 0 : 18,
 821                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 822                                         csv_sep);
 823
 824                                 fprintf(output, "%-*s%s",
 825                                         csv_output ? 0 : unit_width,
 826                                         counter->unit, csv_sep);
 827
 828                                 fprintf(output, "%*s",
 829                                         csv_output ? 0 : -25,
 830                                         perf_evsel__name(counter));
 831
 832                                 if (counter->cgrp)
 833                                         fprintf(output, "%s%s",
 834                                                 csv_sep, counter->cgrp->name);
 835
 836                                 print_running(run, ena);
 837                                 fputc('\n', output);
 838                                 continue;
 839                         }
 840                         uval = val * counter->scale;
 841
 842                         if (nsec_counter(counter))
 843                                 nsec_printout(id, nr, counter, uval);
 844                         else
 845                                 abs_printout(id, nr, counter, uval);
 846
 847                         if (!csv_output)
 848                                 print_noise(counter, 1.0);
 849
 850                         print_running(run, ena);
 851                         fputc('\n', output);
 852                 }
 853         }
 854 }
 855
 856 /*
 857  * Print out the results of a single counter:
 858  * aggregated counts in system-wide mode
 859  */
 860 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
 861 {
 862         struct perf_stat *ps = counter->priv;
 863         double avg = avg_stats(&ps->res_stats[0]);
 864         int scaled = counter->counts->scaled;
 865         double uval;
 866         double avg_enabled, avg_running;
 867
 868         avg_enabled = avg_stats(&ps->res_stats[1]);
 869         avg_running = avg_stats(&ps->res_stats[2]);
 870
 871         if (prefix)
 872                 fprintf(output, "%s", prefix);
 873
 874         if (scaled == -1 || !counter->supported) {
 875                 fprintf(output, "%*s%s",
 876                         csv_output ? 0 : 18,
 877                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 878                         csv_sep);
 879                 fprintf(output, "%-*s%s",
 880                         csv_output ? 0 : unit_width,
 881                         counter->unit, csv_sep);
 882                 fprintf(output, "%*s",
 883                         csv_output ? 0 : -25,
 884                         perf_evsel__name(counter));
 885
 886                 if (counter->cgrp)
 887                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
 888
 889                 print_running(avg_running, avg_enabled);
 890                 fputc('\n', output);
 891                 return;
 892         }
 893
 894         uval = avg * counter->scale;
 895
 896         if (nsec_counter(counter))
 897                 nsec_printout(-1, 0, counter, uval);
 898         else
 899                 abs_printout(-1, 0, counter, uval);
 900
 901         print_noise(counter, avg);
 902
 903         print_running(avg_running, avg_enabled);
 904         fprintf(output, "\n");
 905 }
 906
 907 /*
 908  * Print out the results of a single counter:
 909  * does not use aggregated count in system-wide
 910  */
 911 static void print_counter(struct perf_evsel *counter, char *prefix)
 912 {
 913         u64 ena, run, val;
 914         double uval;
 915         int cpu;
 916
 917         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 918                 val = perf_counts(counter->counts, cpu)->val;
 919                 ena = perf_counts(counter->counts, cpu)->ena;
 920                 run = perf_counts(counter->counts, cpu)->run;
 921
 922                 if (prefix)
 923                         fprintf(output, "%s", prefix);
 924
 925                 if (run == 0 || ena == 0) {
 926                         fprintf(output, "CPU%*d%s%*s%s",
 927                                 csv_output ? 0 : -4,
 928                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
 929                                 csv_output ? 0 : 18,
 930                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 931                                 csv_sep);
 932
 933                                 fprintf(output, "%-*s%s",
 934                                         csv_output ? 0 : unit_width,
 935                                         counter->unit, csv_sep);
 936
 937                                 fprintf(output, "%*s",
 938                                         csv_output ? 0 : -25,
 939                                         perf_evsel__name(counter));
 940
 941                         if (counter->cgrp)
 942                                 fprintf(output, "%s%s",
 943                                         csv_sep, counter->cgrp->name);
 944
 945                         print_running(run, ena);
 946                         fputc('\n', output);
 947                         continue;
 948                 }
 949
 950                 uval = val * counter->scale;
 951
 952                 if (nsec_counter(counter))
 953                         nsec_printout(cpu, 0, counter, uval);
 954                 else
 955                         abs_printout(cpu, 0, counter, uval);
 956
 957                 if (!csv_output)
 958                         print_noise(counter, 1.0);
 959                 print_running(run, ena);
 960
 961                 fputc('\n', output);
 962         }
 963 }
 964
 965 static void print_stat(int argc, const char **argv)
 966 {
 967         struct perf_evsel *counter;
 968         int i;
 969
 970         fflush(stdout);
 971
 972         if (!csv_output) {
 973                 fprintf(output, "\n");
 974                 fprintf(output, " Performance counter stats for ");
 975                 if (target.system_wide)
 976                         fprintf(output, "\'system wide");
 977                 else if (target.cpu_list)
 978                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
 979                 else if (!target__has_task(&target)) {
 980                         fprintf(output, "\'%s", argv[0]);
 981                         for (i = 1; i < argc; i++)
 982                                 fprintf(output, " %s", argv[i]);
 983                 } else if (target.pid)
 984                         fprintf(output, "process id \'%s", target.pid);
 985                 else
 986                         fprintf(output, "thread id \'%s", target.tid);
 987
 988                 fprintf(output, "\'");
 989                 if (run_count > 1)
 990                         fprintf(output, " (%d runs)", run_count);
 991                 fprintf(output, ":\n\n");
 992         }
 993
 994         switch (aggr_mode) {
 995         case AGGR_CORE:
 996         case AGGR_SOCKET:
 997                 print_aggr(NULL);
 998                 break;
 999         case AGGR_GLOBAL:
1000                 evlist__for_each(evsel_list, counter)
1001                         print_counter_aggr(counter, NULL);
1002                 break;
1003         case AGGR_NONE:
1004                 evlist__for_each(evsel_list, counter)
1005                         print_counter(counter, NULL);
1006                 break;
1007         default:
1008                 break;
1009         }
1010
1011         if (!csv_output) {
1012                 if (!null_run)
1013                         fprintf(output, "\n");
1014                 fprintf(output, " %17.9f seconds time elapsed",
1015                                 avg_stats(&walltime_nsecs_stats)/1e9);
1016                 if (run_count > 1) {
1017                         fprintf(output, "                                        ");
1018                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1019                                         avg_stats(&walltime_nsecs_stats));
1020                 }
1021                 fprintf(output, "\n\n");
1022         }
1023 }
1024
1025 static volatile int signr = -1;
1026
1027 static void skip_signal(int signo)
1028 {
1029         if ((child_pid == -1) || interval)
1030                 done = 1;
1031
1032         signr = signo;
1033         /*
1034          * render child_pid harmless
1035          * won't send SIGTERM to a random
1036          * process in case of race condition
1037          * and fast PID recycling
1038          */
1039         child_pid = -1;
1040 }
1041
1042 static void sig_atexit(void)
1043 {
1044         sigset_t set, oset;
1045
1046         /*
1047          * avoid race condition with SIGCHLD handler
1048          * in skip_signal() which is modifying child_pid
1049          * goal is to avoid send SIGTERM to a random
1050          * process
1051          */
1052         sigemptyset(&set);
1053         sigaddset(&set, SIGCHLD);
1054         sigprocmask(SIG_BLOCK, &set, &oset);
1055
1056         if (child_pid != -1)
1057                 kill(child_pid, SIGTERM);
1058
1059         sigprocmask(SIG_SETMASK, &oset, NULL);
1060
1061         if (signr == -1)
1062                 return;
1063
1064         signal(signr, SIG_DFL);
1065         kill(getpid(), signr);
1066 }
1067
1068 static int stat__set_big_num(const struct option *opt __maybe_unused,
1069                              const char *s __maybe_unused, int unset)
1070 {
1071         big_num_opt = unset ? 0 : 1;
1072         return 0;
1073 }
1074
1075 static int perf_stat_init_aggr_mode(void)
1076 {
1077         switch (aggr_mode) {
1078         case AGGR_SOCKET:
1079                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1080                         perror("cannot build socket map");
1081                         return -1;
1082                 }
1083                 aggr_get_id = cpu_map__get_socket;
1084                 break;
1085         case AGGR_CORE:
1086                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1087                         perror("cannot build core map");
1088                         return -1;
1089                 }
1090                 aggr_get_id = cpu_map__get_core;
1091                 break;
1092         case AGGR_NONE:
1093         case AGGR_GLOBAL:
1094         default:
1095                 break;
1096         }
1097         return 0;
1098 }
1099
1100 /*
1101  * Add default attributes, if there were no attributes specified or
1102  * if -d/--detailed, -d -d or -d -d -d is used:
1103  */
1104 static int add_default_attributes(void)
1105 {
1106         struct perf_event_attr default_attrs[] = {
1107
1108   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1109   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1110   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1111   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1112
1113   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1114   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1115   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1116   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1117   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1118   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1119
1120 };
1121
1122 /*
1123  * Detailed stats (-d), covering the L1 and last level data caches:
1124  */
1125         struct perf_event_attr detailed_attrs[] = {
1126
1127   { .type = PERF_TYPE_HW_CACHE,
1128     .config =
1129          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1130         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1131         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1132
1133   { .type = PERF_TYPE_HW_CACHE,
1134     .config =
1135          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1136         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1137         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1138
1139   { .type = PERF_TYPE_HW_CACHE,
1140     .config =
1141          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1142         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1143         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1144
1145   { .type = PERF_TYPE_HW_CACHE,
1146     .config =
1147          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1148         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1149         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1150 };
1151
1152 /*
1153  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1154  */
1155         struct perf_event_attr very_detailed_attrs[] = {
1156
1157   { .type = PERF_TYPE_HW_CACHE,
1158     .config =
1159          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1160         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1161         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1162
1163   { .type = PERF_TYPE_HW_CACHE,
1164     .config =
1165          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1166         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1167         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1168
1169   { .type = PERF_TYPE_HW_CACHE,
1170     .config =
1171          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1172         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1173         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1174
1175   { .type = PERF_TYPE_HW_CACHE,
1176     .config =
1177          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1178         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1179         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1180
1181   { .type = PERF_TYPE_HW_CACHE,
1182     .config =
1183          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1184         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1185         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1186
1187   { .type = PERF_TYPE_HW_CACHE,
1188     .config =
1189          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1190         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1191         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1192
1193 };
1194
1195 /*
1196  * Very, very detailed stats (-d -d -d), adding prefetch events:
1197  */
1198         struct perf_event_attr very_very_detailed_attrs[] = {
1199
1200   { .type = PERF_TYPE_HW_CACHE,
1201     .config =
1202          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1203         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1204         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1205
1206   { .type = PERF_TYPE_HW_CACHE,
1207     .config =
1208          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1209         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1210         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1211 };
1212
1213         /* Set attrs if no event is selected and !null_run: */
1214         if (null_run)
1215                 return 0;
1216
1217         if (transaction_run) {
1218                 int err;
1219                 if (pmu_have_event("cpu", "cycles-ct") &&
1220                     pmu_have_event("cpu", "el-start"))
1221                         err = parse_events(evsel_list, transaction_attrs, NULL);
1222                 else
1223                         err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1224                 if (err) {
1225                         fprintf(stderr, "Cannot set up transaction events\n");
1226                         return -1;
1227                 }
1228                 return 0;
1229         }
1230
1231         if (!evsel_list->nr_entries) {
1232                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1233                         return -1;
1234         }
1235
1236         /* Detailed events get appended to the event list: */
1237
1238         if (detailed_run <  1)
1239                 return 0;
1240
1241         /* Append detailed run extra attributes: */
1242         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1243                 return -1;
1244
1245         if (detailed_run < 2)
1246                 return 0;
1247
1248         /* Append very detailed run extra attributes: */
1249         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1250                 return -1;
1251
1252         if (detailed_run < 3)
1253                 return 0;
1254
1255         /* Append very, very detailed run extra attributes: */
1256         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1257 }
1258
1259 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1260 {
1261         bool append_file = false;
1262         int output_fd = 0;
1263         const char *output_name = NULL;
1264         const struct option options[] = {
1265         OPT_BOOLEAN('T', "transaction", &transaction_run,
1266                     "hardware transaction statistics"),
1267         OPT_CALLBACK('e', "event", &evsel_list, "event",
1268                      "event selector. use 'perf list' to list available events",
1269                      parse_events_option),
1270         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1271                      "event filter", parse_filter),
1272         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1273                     "child tasks do not inherit counters"),
1274         OPT_STRING('p', "pid", &target.pid, "pid",
1275                    "stat events on existing process id"),
1276         OPT_STRING('t', "tid", &target.tid, "tid",
1277                    "stat events on existing thread id"),
1278         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1279                     "system-wide collection from all CPUs"),
1280         OPT_BOOLEAN('g', "group", &group,
1281                     "put the counters into a counter group"),
1282         OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1283         OPT_INCR('v', "verbose", &verbose,
1284                     "be more verbose (show counter open errors, etc)"),
1285         OPT_INTEGER('r', "repeat", &run_count,
1286                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1287         OPT_BOOLEAN('n', "null", &null_run,
1288                     "null run - dont start any counters"),
1289         OPT_INCR('d', "detailed", &detailed_run,
1290                     "detailed run - start a lot of events"),
1291         OPT_BOOLEAN('S', "sync", &sync_run,
1292                     "call sync() before starting a run"),
1293         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1294                            "print large numbers with thousands\' separators",
1295                            stat__set_big_num),
1296         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1297                     "list of cpus to monitor in system-wide"),
1298         OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1299                     "disable CPU count aggregation", AGGR_NONE),
1300         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1301                    "print counts with custom separator"),
1302         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1303                      "monitor event in cgroup name only", parse_cgroups),
1304         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1305         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1306         OPT_INTEGER(0, "log-fd", &output_fd,
1307                     "log output to fd, instead of stderr"),
1308         OPT_STRING(0, "pre", &pre_cmd, "command",
1309                         "command to run prior to the measured command"),
1310         OPT_STRING(0, "post", &post_cmd, "command",
1311                         "command to run after to the measured command"),
1312         OPT_UINTEGER('I', "interval-print", &interval,
1313                     "print counts at regular interval in ms (>= 100)"),
1314         OPT_SET_UINT(0, "per-socket", &aggr_mode,
1315                      "aggregate counts per processor socket", AGGR_SOCKET),
1316         OPT_SET_UINT(0, "per-core", &aggr_mode,
1317                      "aggregate counts per physical processor core", AGGR_CORE),
1318         OPT_UINTEGER('D', "delay", &initial_delay,
1319                      "ms to wait before starting measurement after program start"),
1320         OPT_END()
1321         };
1322         const char * const stat_usage[] = {
1323                 "perf stat [<options>] [<command>]",
1324                 NULL
1325         };
1326         int status = -EINVAL, run_idx;
1327         const char *mode;
1328
1329         setlocale(LC_ALL, "");
1330
1331         evsel_list = perf_evlist__new();
1332         if (evsel_list == NULL)
1333                 return -ENOMEM;
1334
1335         argc = parse_options(argc, argv, options, stat_usage,
1336                 PARSE_OPT_STOP_AT_NON_OPTION);
1337
1338         output = stderr;
1339         if (output_name && strcmp(output_name, "-"))
1340                 output = NULL;
1341
1342         if (output_name && output_fd) {
1343                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1344                 parse_options_usage(stat_usage, options, "o", 1);
1345                 parse_options_usage(NULL, options, "log-fd", 0);
1346                 goto out;
1347         }
1348
1349         if (output_fd < 0) {
1350                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1351                 parse_options_usage(stat_usage, options, "log-fd", 0);
1352                 goto out;
1353         }
1354
1355         if (!output) {
1356                 struct timespec tm;
1357                 mode = append_file ? "a" : "w";
1358
1359                 output = fopen(output_name, mode);
1360                 if (!output) {
1361                         perror("failed to create output file");
1362                         return -1;
1363                 }
1364                 clock_gettime(CLOCK_REALTIME, &tm);
1365                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1366         } else if (output_fd > 0) {
1367                 mode = append_file ? "a" : "w";
1368                 output = fdopen(output_fd, mode);
1369                 if (!output) {
1370                         perror("Failed opening logfd");
1371                         return -errno;
1372                 }
1373         }
1374
1375         if (csv_sep) {
1376                 csv_output = true;
1377                 if (!strcmp(csv_sep, "\\t"))
1378                         csv_sep = "\t";
1379         } else
1380                 csv_sep = DEFAULT_SEPARATOR;
1381
1382         /*
1383          * let the spreadsheet do the pretty-printing
1384          */
1385         if (csv_output) {
1386                 /* User explicitly passed -B? */
1387                 if (big_num_opt == 1) {
1388                         fprintf(stderr, "-B option not supported with -x\n");
1389                         parse_options_usage(stat_usage, options, "B", 1);
1390                         parse_options_usage(NULL, options, "x", 1);
1391                         goto out;
1392                 } else /* Nope, so disable big number formatting */
1393                         big_num = false;
1394         } else if (big_num_opt == 0) /* User passed --no-big-num */
1395                 big_num = false;
1396
1397         if (!argc && target__none(&target))
1398                 usage_with_options(stat_usage, options);
1399
1400         if (run_count < 0) {
1401                 pr_err("Run count must be a positive number\n");
1402                 parse_options_usage(stat_usage, options, "r", 1);
1403                 goto out;
1404         } else if (run_count == 0) {
1405                 forever = true;
1406                 run_count = 1;
1407         }
1408
1409         /* no_aggr, cgroup are for system-wide only */
1410         if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1411             !target__has_cpu(&target)) {
1412                 fprintf(stderr, "both cgroup and no-aggregation "
1413                         "modes only available in system-wide mode\n");
1414
1415                 parse_options_usage(stat_usage, options, "G", 1);
1416                 parse_options_usage(NULL, options, "A", 1);
1417                 parse_options_usage(NULL, options, "a", 1);
1418                 goto out;
1419         }
1420
1421         if (add_default_attributes())
1422                 goto out;
1423
1424         target__validate(&target);
1425
1426         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1427                 if (target__has_task(&target)) {
1428                         pr_err("Problems finding threads of monitor\n");
1429                         parse_options_usage(stat_usage, options, "p", 1);
1430                         parse_options_usage(NULL, options, "t", 1);
1431                 } else if (target__has_cpu(&target)) {
1432                         perror("failed to parse CPUs map");
1433                         parse_options_usage(stat_usage, options, "C", 1);
1434                         parse_options_usage(NULL, options, "a", 1);
1435                 }
1436                 goto out;
1437         }
1438         if (interval && interval < 100) {
1439                 pr_err("print interval must be >= 100ms\n");
1440                 parse_options_usage(stat_usage, options, "I", 1);
1441                 goto out;
1442         }
1443
1444         if (perf_evlist__alloc_stats(evsel_list, interval))
1445                 goto out;
1446
1447         if (perf_stat_init_aggr_mode())
1448                 goto out;
1449
1450         /*
1451          * We dont want to block the signals - that would cause
1452          * child tasks to inherit that and Ctrl-C would not work.
1453          * What we want is for Ctrl-C to work in the exec()-ed
1454          * task, but being ignored by perf stat itself:
1455          */
1456         atexit(sig_atexit);
1457         if (!forever)
1458                 signal(SIGINT,  skip_signal);
1459         signal(SIGCHLD, skip_signal);
1460         signal(SIGALRM, skip_signal);
1461         signal(SIGABRT, skip_signal);
1462
1463         status = 0;
1464         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1465                 if (run_count != 1 && verbose)
1466                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1467                                 run_idx + 1);
1468
1469                 status = run_perf_stat(argc, argv);
1470                 if (forever && status != -1) {
1471                         print_stat(argc, argv);
1472                         perf_stat__reset_stats(evsel_list);
1473                 }
1474         }
1475
1476         if (!forever && status != -1 && !interval)
1477                 print_stat(argc, argv);
1478
1479         perf_evlist__free_stats(evsel_list);
1480 out:
1481         perf_evlist__delete(evsel_list);
1482         return status;
1483 }