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    Time: 0.104
  11
  12     Performance counter stats for '/home/mingo/hackbench':
  13
  14        1255.538611  task clock ticks     #      10.143 CPU utilization factor
  15              54011  context switches     #       0.043 M/sec
  16                385  CPU migrations       #       0.000 M/sec
  17              17755  pagefaults           #       0.014 M/sec
  18         3808323185  CPU cycles           #    3033.219 M/sec
  19         1575111190  instructions         #    1254.530 M/sec
  20           17367895  cache references     #      13.833 M/sec
  21            7674421  cache misses         #       6.112 M/sec
  22
  23     Wall-clock time elapsed:   123.786620 msecs
  24
  25  *
  26  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  27  *
  28  * Improvements and fixes by:
  29  *
  30  *   Arjan van de Ven <arjan@linux.intel.com>
  31  *   Yanmin Zhang <yanmin.zhang@intel.com>
  32  *   Wu Fengguang <fengguang.wu@intel.com>
  33  *   Mike Galbraith <efault@gmx.de>
  34  *   Paul Mackerras <paulus@samba.org>
  35  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  36  *
  37  * Released under the GPL v2. (and only v2, not any later version)
  38  */
  39
  40 #include "perf.h"
  41 #include "builtin.h"
  42 #include "util/util.h"
  43 #include "util/parse-options.h"
  44 #include "util/parse-events.h"
  45 #include "util/event.h"
  46 #include "util/evlist.h"
  47 #include "util/evsel.h"
  48 #include "util/debug.h"
  49 #include "util/header.h"
  50 #include "util/cpumap.h"
  51 #include "util/thread.h"
  52 #include "util/thread_map.h"
  53
  54 #include <sys/prctl.h>
  55 #include <math.h>
  56 #include <locale.h>
  57
  58 #define DEFAULT_SEPARATOR       " "
  59
  60 static struct perf_event_attr default_attrs[] = {
  61
  62   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
  63   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
  64   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
  65   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
  66
  67   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
  68   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
  69   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
  70   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
  71   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES        },
  72   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES            },
  73
  74 };
  75
  76 struct perf_evlist              *evsel_list;
  77
  78 static bool                     system_wide                     =  false;
  79 static int                      run_idx                         =  0;
  80
  81 static int                      run_count                       =  1;
  82 static bool                     no_inherit                      = false;
  83 static bool                     scale                           =  true;
  84 static bool                     no_aggr                         = false;
  85 static pid_t                    target_pid                      = -1;
  86 static pid_t                    target_tid                      = -1;
  87 static pid_t                    child_pid                       = -1;
  88 static bool                     null_run                        =  false;
  89 static bool                     big_num                         =  true;
  90 static int                      big_num_opt                     =  -1;
  91 static const char               *cpu_list;
  92 static const char               *csv_sep                        = NULL;
  93 static bool                     csv_output                      = false;
  94
  95 static volatile int done = 0;
  96
  97 struct stats
  98 {
  99         double n, mean, M2;
 100 };
 101
 102 struct perf_stat {
 103         struct stats      res_stats[3];
 104 };
 105
 106 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 107 {
 108         evsel->priv = zalloc(sizeof(struct perf_stat));
 109         return evsel->priv == NULL ? -ENOMEM : 0;
 110 }
 111
 112 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 113 {
 114         free(evsel->priv);
 115         evsel->priv = NULL;
 116 }
 117
 118 static void update_stats(struct stats *stats, u64 val)
 119 {
 120         double delta;
 121
 122         stats->n++;
 123         delta = val - stats->mean;
 124         stats->mean += delta / stats->n;
 125         stats->M2 += delta*(val - stats->mean);
 126 }
 127
 128 static double avg_stats(struct stats *stats)
 129 {
 130         return stats->mean;
 131 }
 132
 133 /*
 134  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 135  *
 136  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 137  * s^2 = -------------------------------
 138  *                  n - 1
 139  *
 140  * http://en.wikipedia.org/wiki/Stddev
 141  *
 142  * The std dev of the mean is related to the std dev by:
 143  *
 144  *             s
 145  * s_mean = -------
 146  *          sqrt(n)
 147  *
 148  */
 149 static double stddev_stats(struct stats *stats)
 150 {
 151         double variance = stats->M2 / (stats->n - 1);
 152         double variance_mean = variance / stats->n;
 153
 154         return sqrt(variance_mean);
 155 }
 156
 157 struct stats                    runtime_nsecs_stats[MAX_NR_CPUS];
 158 struct stats                    runtime_cycles_stats[MAX_NR_CPUS];
 159 struct stats                    runtime_branches_stats[MAX_NR_CPUS];
 160 struct stats                    walltime_nsecs_stats;
 161
 162 static int create_perf_stat_counter(struct perf_evsel *evsel)
 163 {
 164         struct perf_event_attr *attr = &evsel->attr;
 165
 166         if (scale)
 167                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 168                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 169
 170         if (system_wide)
 171                 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false, false);
 172
 173         attr->inherit = !no_inherit;
 174         if (target_pid == -1 && target_tid == -1) {
 175                 attr->disabled = 1;
 176                 attr->enable_on_exec = 1;
 177         }
 178
 179         return perf_evsel__open_per_thread(evsel, evsel_list->threads, false, false);
 180 }
 181
 182 /*
 183  * Does the counter have nsecs as a unit?
 184  */
 185 static inline int nsec_counter(struct perf_evsel *evsel)
 186 {
 187         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 188             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 189                 return 1;
 190
 191         return 0;
 192 }
 193
 194 /*
 195  * Read out the results of a single counter:
 196  * aggregate counts across CPUs in system-wide mode
 197  */
 198 static int read_counter_aggr(struct perf_evsel *counter)
 199 {
 200         struct perf_stat *ps = counter->priv;
 201         u64 *count = counter->counts->aggr.values;
 202         int i;
 203
 204         if (__perf_evsel__read(counter, evsel_list->cpus->nr,
 205                                evsel_list->threads->nr, scale) < 0)
 206                 return -1;
 207
 208         for (i = 0; i < 3; i++)
 209                 update_stats(&ps->res_stats[i], count[i]);
 210
 211         if (verbose) {
 212                 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 213                         event_name(counter), count[0], count[1], count[2]);
 214         }
 215
 216         /*
 217          * Save the full runtime - to allow normalization during printout:
 218          */
 219         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 220                 update_stats(&runtime_nsecs_stats[0], count[0]);
 221         if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 222                 update_stats(&runtime_cycles_stats[0], count[0]);
 223         if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 224                 update_stats(&runtime_branches_stats[0], count[0]);
 225
 226         return 0;
 227 }
 228
 229 /*
 230  * Read out the results of a single counter:
 231  * do not aggregate counts across CPUs in system-wide mode
 232  */
 233 static int read_counter(struct perf_evsel *counter)
 234 {
 235         u64 *count;
 236         int cpu;
 237
 238         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 239                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 240                         return -1;
 241
 242                 count = counter->counts->cpu[cpu].values;
 243
 244                 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 245                         update_stats(&runtime_nsecs_stats[cpu], count[0]);
 246                 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 247                         update_stats(&runtime_cycles_stats[cpu], count[0]);
 248                 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 249                         update_stats(&runtime_branches_stats[cpu], count[0]);
 250         }
 251
 252         return 0;
 253 }
 254
 255 static int run_perf_stat(int argc __used, const char **argv)
 256 {
 257         unsigned long long t0, t1;
 258         struct perf_evsel *counter;
 259         int status = 0;
 260         int child_ready_pipe[2], go_pipe[2];
 261         const bool forks = (argc > 0);
 262         char buf;
 263
 264         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 265                 perror("failed to create pipes");
 266                 exit(1);
 267         }
 268
 269         if (forks) {
 270                 if ((child_pid = fork()) < 0)
 271                         perror("failed to fork");
 272
 273                 if (!child_pid) {
 274                         close(child_ready_pipe[0]);
 275                         close(go_pipe[1]);
 276                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 277
 278                         /*
 279                          * Do a dummy execvp to get the PLT entry resolved,
 280                          * so we avoid the resolver overhead on the real
 281                          * execvp call.
 282                          */
 283                         execvp("", (char **)argv);
 284
 285                         /*
 286                          * Tell the parent we're ready to go
 287                          */
 288                         close(child_ready_pipe[1]);
 289
 290                         /*
 291                          * Wait until the parent tells us to go.
 292                          */
 293                         if (read(go_pipe[0], &buf, 1) == -1)
 294                                 perror("unable to read pipe");
 295
 296                         execvp(argv[0], (char **)argv);
 297
 298                         perror(argv[0]);
 299                         exit(-1);
 300                 }
 301
 302                 if (target_tid == -1 && target_pid == -1 && !system_wide)
 303                         evsel_list->threads->map[0] = child_pid;
 304
 305                 /*
 306                  * Wait for the child to be ready to exec.
 307                  */
 308                 close(child_ready_pipe[1]);
 309                 close(go_pipe[0]);
 310                 if (read(child_ready_pipe[0], &buf, 1) == -1)
 311                         perror("unable to read pipe");
 312                 close(child_ready_pipe[0]);
 313         }
 314
 315         list_for_each_entry(counter, &evsel_list->entries, node) {
 316                 if (create_perf_stat_counter(counter) < 0) {
 317                         if (errno == -EPERM || errno == -EACCES) {
 318                                 error("You may not have permission to collect %sstats.\n"
 319                                       "\t Consider tweaking"
 320                                       " /proc/sys/kernel/perf_event_paranoid or running as root.",
 321                                       system_wide ? "system-wide " : "");
 322                         } else if (errno == ENOENT) {
 323                                 error("%s event is not supported. ", event_name(counter));
 324                         } else {
 325                                 error("open_counter returned with %d (%s). "
 326                                       "/bin/dmesg may provide additional information.\n",
 327                                        errno, strerror(errno));
 328                         }
 329                         if (child_pid != -1)
 330                                 kill(child_pid, SIGTERM);
 331                         die("Not all events could be opened.\n");
 332                         return -1;
 333                 }
 334         }
 335
 336         if (perf_evlist__set_filters(evsel_list)) {
 337                 error("failed to set filter with %d (%s)\n", errno,
 338                         strerror(errno));
 339                 return -1;
 340         }
 341
 342         /*
 343          * Enable counters and exec the command:
 344          */
 345         t0 = rdclock();
 346
 347         if (forks) {
 348                 close(go_pipe[1]);
 349                 wait(&status);
 350         } else {
 351                 while(!done) sleep(1);
 352         }
 353
 354         t1 = rdclock();
 355
 356         update_stats(&walltime_nsecs_stats, t1 - t0);
 357
 358         if (no_aggr) {
 359                 list_for_each_entry(counter, &evsel_list->entries, node) {
 360                         read_counter(counter);
 361                         perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
 362                 }
 363         } else {
 364                 list_for_each_entry(counter, &evsel_list->entries, node) {
 365                         read_counter_aggr(counter);
 366                         perf_evsel__close_fd(counter, evsel_list->cpus->nr,
 367                                              evsel_list->threads->nr);
 368                 }
 369         }
 370
 371         return WEXITSTATUS(status);
 372 }
 373
 374 static void print_noise(struct perf_evsel *evsel, double avg)
 375 {
 376         struct perf_stat *ps;
 377
 378         if (run_count == 1)
 379                 return;
 380
 381         ps = evsel->priv;
 382         fprintf(stderr, "   ( +- %7.3f%% )",
 383                         100 * stddev_stats(&ps->res_stats[0]) / avg);
 384 }
 385
 386 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 387 {
 388         double msecs = avg / 1e6;
 389         char cpustr[16] = { '\0', };
 390         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
 391
 392         if (no_aggr)
 393                 sprintf(cpustr, "CPU%*d%s",
 394                         csv_output ? 0 : -4,
 395                         evsel_list->cpus->map[cpu], csv_sep);
 396
 397         fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
 398
 399         if (evsel->cgrp)
 400                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 401
 402         if (csv_output)
 403                 return;
 404
 405         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 406                 fprintf(stderr, " # %10.3f CPUs ",
 407                                 avg / avg_stats(&walltime_nsecs_stats));
 408 }
 409
 410 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 411 {
 412         double total, ratio = 0.0;
 413         char cpustr[16] = { '\0', };
 414         const char *fmt;
 415
 416         if (csv_output)
 417                 fmt = "%s%.0f%s%s";
 418         else if (big_num)
 419                 fmt = "%s%'18.0f%s%-24s";
 420         else
 421                 fmt = "%s%18.0f%s%-24s";
 422
 423         if (no_aggr)
 424                 sprintf(cpustr, "CPU%*d%s",
 425                         csv_output ? 0 : -4,
 426                         evsel_list->cpus->map[cpu], csv_sep);
 427         else
 428                 cpu = 0;
 429
 430         fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
 431
 432         if (evsel->cgrp)
 433                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 434
 435         if (csv_output)
 436                 return;
 437
 438         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 439                 total = avg_stats(&runtime_cycles_stats[cpu]);
 440
 441                 if (total)
 442                         ratio = avg / total;
 443
 444                 fprintf(stderr, " # %10.3f IPC  ", ratio);
 445         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 446                         runtime_branches_stats[cpu].n != 0) {
 447                 total = avg_stats(&runtime_branches_stats[cpu]);
 448
 449                 if (total)
 450                         ratio = avg * 100 / total;
 451
 452                 fprintf(stderr, " # %10.3f %%    ", ratio);
 453
 454         } else if (runtime_nsecs_stats[cpu].n != 0) {
 455                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 456
 457                 if (total)
 458                         ratio = 1000.0 * avg / total;
 459
 460                 fprintf(stderr, " # %10.3f M/sec", ratio);
 461         }
 462 }
 463
 464 /*
 465  * Print out the results of a single counter:
 466  * aggregated counts in system-wide mode
 467  */
 468 static void print_counter_aggr(struct perf_evsel *counter)
 469 {
 470         struct perf_stat *ps = counter->priv;
 471         double avg = avg_stats(&ps->res_stats[0]);
 472         int scaled = counter->counts->scaled;
 473
 474         if (scaled == -1) {
 475                 fprintf(stderr, "%*s%s%*s",
 476                         csv_output ? 0 : 18,
 477                         "<not counted>",
 478                         csv_sep,
 479                         csv_output ? 0 : -24,
 480                         event_name(counter));
 481
 482                 if (counter->cgrp)
 483                         fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 484
 485                 fputc('\n', stderr);
 486                 return;
 487         }
 488
 489         if (nsec_counter(counter))
 490                 nsec_printout(-1, counter, avg);
 491         else
 492                 abs_printout(-1, counter, avg);
 493
 494         if (csv_output) {
 495                 fputc('\n', stderr);
 496                 return;
 497         }
 498
 499         print_noise(counter, avg);
 500
 501         if (scaled) {
 502                 double avg_enabled, avg_running;
 503
 504                 avg_enabled = avg_stats(&ps->res_stats[1]);
 505                 avg_running = avg_stats(&ps->res_stats[2]);
 506
 507                 fprintf(stderr, "  (scaled from %.2f%%)",
 508                                 100 * avg_running / avg_enabled);
 509         }
 510         fprintf(stderr, "\n");
 511 }
 512
 513 /*
 514  * Print out the results of a single counter:
 515  * does not use aggregated count in system-wide
 516  */
 517 static void print_counter(struct perf_evsel *counter)
 518 {
 519         u64 ena, run, val;
 520         int cpu;
 521
 522         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 523                 val = counter->counts->cpu[cpu].val;
 524                 ena = counter->counts->cpu[cpu].ena;
 525                 run = counter->counts->cpu[cpu].run;
 526                 if (run == 0 || ena == 0) {
 527                         fprintf(stderr, "CPU%*d%s%*s%s%*s",
 528                                 csv_output ? 0 : -4,
 529                                 evsel_list->cpus->map[cpu], csv_sep,
 530                                 csv_output ? 0 : 18,
 531                                 "<not counted>", csv_sep,
 532                                 csv_output ? 0 : -24,
 533                                 event_name(counter));
 534
 535                         if (counter->cgrp)
 536                                 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 537
 538                         fputc('\n', stderr);
 539                         continue;
 540                 }
 541
 542                 if (nsec_counter(counter))
 543                         nsec_printout(cpu, counter, val);
 544                 else
 545                         abs_printout(cpu, counter, val);
 546
 547                 if (!csv_output) {
 548                         print_noise(counter, 1.0);
 549
 550                         if (run != ena) {
 551                                 fprintf(stderr, "  (scaled from %.2f%%)",
 552                                         100.0 * run / ena);
 553                         }
 554                 }
 555                 fputc('\n', stderr);
 556         }
 557 }
 558
 559 static void print_stat(int argc, const char **argv)
 560 {
 561         struct perf_evsel *counter;
 562         int i;
 563
 564         fflush(stdout);
 565
 566         if (!csv_output) {
 567                 fprintf(stderr, "\n");
 568                 fprintf(stderr, " Performance counter stats for ");
 569                 if(target_pid == -1 && target_tid == -1) {
 570                         fprintf(stderr, "\'%s", argv[0]);
 571                         for (i = 1; i < argc; i++)
 572                                 fprintf(stderr, " %s", argv[i]);
 573                 } else if (target_pid != -1)
 574                         fprintf(stderr, "process id \'%d", target_pid);
 575                 else
 576                         fprintf(stderr, "thread id \'%d", target_tid);
 577
 578                 fprintf(stderr, "\'");
 579                 if (run_count > 1)
 580                         fprintf(stderr, " (%d runs)", run_count);
 581                 fprintf(stderr, ":\n\n");
 582         }
 583
 584         if (no_aggr) {
 585                 list_for_each_entry(counter, &evsel_list->entries, node)
 586                         print_counter(counter);
 587         } else {
 588                 list_for_each_entry(counter, &evsel_list->entries, node)
 589                         print_counter_aggr(counter);
 590         }
 591
 592         if (!csv_output) {
 593                 fprintf(stderr, "\n");
 594                 fprintf(stderr, " %18.9f  seconds time elapsed",
 595                                 avg_stats(&walltime_nsecs_stats)/1e9);
 596                 if (run_count > 1) {
 597                         fprintf(stderr, "   ( +- %7.3f%% )",
 598                                 100*stddev_stats(&walltime_nsecs_stats) /
 599                                 avg_stats(&walltime_nsecs_stats));
 600                 }
 601                 fprintf(stderr, "\n\n");
 602         }
 603 }
 604
 605 static volatile int signr = -1;
 606
 607 static void skip_signal(int signo)
 608 {
 609         if(child_pid == -1)
 610                 done = 1;
 611
 612         signr = signo;
 613 }
 614
 615 static void sig_atexit(void)
 616 {
 617         if (child_pid != -1)
 618                 kill(child_pid, SIGTERM);
 619
 620         if (signr == -1)
 621                 return;
 622
 623         signal(signr, SIG_DFL);
 624         kill(getpid(), signr);
 625 }
 626
 627 static const char * const stat_usage[] = {
 628         "perf stat [<options>] [<command>]",
 629         NULL
 630 };
 631
 632 static int stat__set_big_num(const struct option *opt __used,
 633                              const char *s __used, int unset)
 634 {
 635         big_num_opt = unset ? 0 : 1;
 636         return 0;
 637 }
 638
 639 static const struct option options[] = {
 640         OPT_CALLBACK('e', "event", &evsel_list, "event",
 641                      "event selector. use 'perf list' to list available events",
 642                      parse_events),
 643         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
 644                      "event filter", parse_filter),
 645         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 646                     "child tasks do not inherit counters"),
 647         OPT_INTEGER('p', "pid", &target_pid,
 648                     "stat events on existing process id"),
 649         OPT_INTEGER('t', "tid", &target_tid,
 650                     "stat events on existing thread id"),
 651         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 652                     "system-wide collection from all CPUs"),
 653         OPT_BOOLEAN('c', "scale", &scale,
 654                     "scale/normalize counters"),
 655         OPT_INCR('v', "verbose", &verbose,
 656                     "be more verbose (show counter open errors, etc)"),
 657         OPT_INTEGER('r', "repeat", &run_count,
 658                     "repeat command and print average + stddev (max: 100)"),
 659         OPT_BOOLEAN('n', "null", &null_run,
 660                     "null run - dont start any counters"),
 661         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
 662                            "print large numbers with thousands\' separators",
 663                            stat__set_big_num),
 664         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 665                     "list of cpus to monitor in system-wide"),
 666         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
 667                     "disable CPU count aggregation"),
 668         OPT_STRING('x', "field-separator", &csv_sep, "separator",
 669                    "print counts with custom separator"),
 670         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 671                      "monitor event in cgroup name only",
 672                      parse_cgroups),
 673         OPT_END()
 674 };
 675
 676 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 677 {
 678         struct perf_evsel *pos;
 679         int status = -ENOMEM;
 680
 681         setlocale(LC_ALL, "");
 682
 683         evsel_list = perf_evlist__new(NULL, NULL);
 684         if (evsel_list == NULL)
 685                 return -ENOMEM;
 686
 687         argc = parse_options(argc, argv, options, stat_usage,
 688                 PARSE_OPT_STOP_AT_NON_OPTION);
 689
 690         if (csv_sep)
 691                 csv_output = true;
 692         else
 693                 csv_sep = DEFAULT_SEPARATOR;
 694
 695         /*
 696          * let the spreadsheet do the pretty-printing
 697          */
 698         if (csv_output) {
 699                 /* User explicitely passed -B? */
 700                 if (big_num_opt == 1) {
 701                         fprintf(stderr, "-B option not supported with -x\n");
 702                         usage_with_options(stat_usage, options);
 703                 } else /* Nope, so disable big number formatting */
 704                         big_num = false;
 705         } else if (big_num_opt == 0) /* User passed --no-big-num */
 706                 big_num = false;
 707
 708         if (!argc && target_pid == -1 && target_tid == -1)
 709                 usage_with_options(stat_usage, options);
 710         if (run_count <= 0)
 711                 usage_with_options(stat_usage, options);
 712
 713         /* no_aggr, cgroup are for system-wide only */
 714         if ((no_aggr || nr_cgroups) && !system_wide) {
 715                 fprintf(stderr, "both cgroup and no-aggregation "
 716                         "modes only available in system-wide mode\n");
 717
 718                 usage_with_options(stat_usage, options);
 719         }
 720
 721         /* Set attrs and nr_counters if no event is selected and !null_run */
 722         if (!null_run && !evsel_list->nr_entries) {
 723                 size_t c;
 724
 725                 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
 726                         pos = perf_evsel__new(&default_attrs[c], c);
 727                         if (pos == NULL)
 728                                 goto out;
 729                         perf_evlist__add(evsel_list, pos);
 730                 }
 731         }
 732
 733         if (target_pid != -1)
 734                 target_tid = target_pid;
 735
 736         evsel_list->threads = thread_map__new(target_pid, target_tid);
 737         if (evsel_list->threads == NULL) {
 738                 pr_err("Problems finding threads of monitor\n");
 739                 usage_with_options(stat_usage, options);
 740         }
 741
 742         if (system_wide)
 743                 evsel_list->cpus = cpu_map__new(cpu_list);
 744         else
 745                 evsel_list->cpus = cpu_map__dummy_new();
 746
 747         if (evsel_list->cpus == NULL) {
 748                 perror("failed to parse CPUs map");
 749                 usage_with_options(stat_usage, options);
 750                 return -1;
 751         }
 752
 753         list_for_each_entry(pos, &evsel_list->entries, node) {
 754                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
 755                     perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
 756                     perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
 757                         goto out_free_fd;
 758         }
 759
 760         /*
 761          * We dont want to block the signals - that would cause
 762          * child tasks to inherit that and Ctrl-C would not work.
 763          * What we want is for Ctrl-C to work in the exec()-ed
 764          * task, but being ignored by perf stat itself:
 765          */
 766         atexit(sig_atexit);
 767         signal(SIGINT,  skip_signal);
 768         signal(SIGALRM, skip_signal);
 769         signal(SIGABRT, skip_signal);
 770
 771         status = 0;
 772         for (run_idx = 0; run_idx < run_count; run_idx++) {
 773                 if (run_count != 1 && verbose)
 774                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 775                 status = run_perf_stat(argc, argv);
 776         }
 777
 778         if (status != -1)
 779                 print_stat(argc, argv);
 780 out_free_fd:
 781         list_for_each_entry(pos, &evsel_list->entries, node)
 782                 perf_evsel__free_stat_priv(pos);
 783         perf_evlist__delete_maps(evsel_list);
 784 out:
 785         perf_evlist__delete(evsel_list);
 786         return status;
 787 }