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         attr->inherit = !no_inherit;
 171
 172         if (system_wide)
 173                 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false);
 174
 175         if (target_pid == -1 && target_tid == -1) {
 176                 attr->disabled = 1;
 177                 attr->enable_on_exec = 1;
 178         }
 179
 180         return perf_evsel__open_per_thread(evsel, evsel_list->threads, false);
 181 }
 182
 183 /*
 184  * Does the counter have nsecs as a unit?
 185  */
 186 static inline int nsec_counter(struct perf_evsel *evsel)
 187 {
 188         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 189             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 190                 return 1;
 191
 192         return 0;
 193 }
 194
 195 /*
 196  * Read out the results of a single counter:
 197  * aggregate counts across CPUs in system-wide mode
 198  */
 199 static int read_counter_aggr(struct perf_evsel *counter)
 200 {
 201         struct perf_stat *ps = counter->priv;
 202         u64 *count = counter->counts->aggr.values;
 203         int i;
 204
 205         if (__perf_evsel__read(counter, evsel_list->cpus->nr,
 206                                evsel_list->threads->nr, scale) < 0)
 207                 return -1;
 208
 209         for (i = 0; i < 3; i++)
 210                 update_stats(&ps->res_stats[i], count[i]);
 211
 212         if (verbose) {
 213                 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 214                         event_name(counter), count[0], count[1], count[2]);
 215         }
 216
 217         /*
 218          * Save the full runtime - to allow normalization during printout:
 219          */
 220         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 221                 update_stats(&runtime_nsecs_stats[0], count[0]);
 222         if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 223                 update_stats(&runtime_cycles_stats[0], count[0]);
 224         if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 225                 update_stats(&runtime_branches_stats[0], count[0]);
 226
 227         return 0;
 228 }
 229
 230 /*
 231  * Read out the results of a single counter:
 232  * do not aggregate counts across CPUs in system-wide mode
 233  */
 234 static int read_counter(struct perf_evsel *counter)
 235 {
 236         u64 *count;
 237         int cpu;
 238
 239         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 240                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 241                         return -1;
 242
 243                 count = counter->counts->cpu[cpu].values;
 244
 245                 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 246                         update_stats(&runtime_nsecs_stats[cpu], count[0]);
 247                 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 248                         update_stats(&runtime_cycles_stats[cpu], count[0]);
 249                 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 250                         update_stats(&runtime_branches_stats[cpu], count[0]);
 251         }
 252
 253         return 0;
 254 }
 255
 256 static int run_perf_stat(int argc __used, const char **argv)
 257 {
 258         unsigned long long t0, t1;
 259         struct perf_evsel *counter;
 260         int status = 0;
 261         int child_ready_pipe[2], go_pipe[2];
 262         const bool forks = (argc > 0);
 263         char buf;
 264
 265         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 266                 perror("failed to create pipes");
 267                 exit(1);
 268         }
 269
 270         if (forks) {
 271                 if ((child_pid = fork()) < 0)
 272                         perror("failed to fork");
 273
 274                 if (!child_pid) {
 275                         close(child_ready_pipe[0]);
 276                         close(go_pipe[1]);
 277                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 278
 279                         /*
 280                          * Do a dummy execvp to get the PLT entry resolved,
 281                          * so we avoid the resolver overhead on the real
 282                          * execvp call.
 283                          */
 284                         execvp("", (char **)argv);
 285
 286                         /*
 287                          * Tell the parent we're ready to go
 288                          */
 289                         close(child_ready_pipe[1]);
 290
 291                         /*
 292                          * Wait until the parent tells us to go.
 293                          */
 294                         if (read(go_pipe[0], &buf, 1) == -1)
 295                                 perror("unable to read pipe");
 296
 297                         execvp(argv[0], (char **)argv);
 298
 299                         perror(argv[0]);
 300                         exit(-1);
 301                 }
 302
 303                 if (target_tid == -1 && target_pid == -1 && !system_wide)
 304                         evsel_list->threads->map[0] = child_pid;
 305
 306                 /*
 307                  * Wait for the child to be ready to exec.
 308                  */
 309                 close(child_ready_pipe[1]);
 310                 close(go_pipe[0]);
 311                 if (read(child_ready_pipe[0], &buf, 1) == -1)
 312                         perror("unable to read pipe");
 313                 close(child_ready_pipe[0]);
 314         }
 315
 316         list_for_each_entry(counter, &evsel_list->entries, node) {
 317                 if (create_perf_stat_counter(counter) < 0) {
 318                         if (errno == -EPERM || errno == -EACCES) {
 319                                 error("You may not have permission to collect %sstats.\n"
 320                                       "\t Consider tweaking"
 321                                       " /proc/sys/kernel/perf_event_paranoid or running as root.",
 322                                       system_wide ? "system-wide " : "");
 323                         } else if (errno == ENOENT) {
 324                                 error("%s event is not supported. ", event_name(counter));
 325                         } else {
 326                                 error("open_counter returned with %d (%s). "
 327                                       "/bin/dmesg may provide additional information.\n",
 328                                        errno, strerror(errno));
 329                         }
 330                         if (child_pid != -1)
 331                                 kill(child_pid, SIGTERM);
 332                         die("Not all events could be opened.\n");
 333                         return -1;
 334                 }
 335         }
 336
 337         if (perf_evlist__set_filters(evsel_list)) {
 338                 error("failed to set filter with %d (%s)\n", errno,
 339                         strerror(errno));
 340                 return -1;
 341         }
 342
 343         /*
 344          * Enable counters and exec the command:
 345          */
 346         t0 = rdclock();
 347
 348         if (forks) {
 349                 close(go_pipe[1]);
 350                 wait(&status);
 351         } else {
 352                 while(!done) sleep(1);
 353         }
 354
 355         t1 = rdclock();
 356
 357         update_stats(&walltime_nsecs_stats, t1 - t0);
 358
 359         if (no_aggr) {
 360                 list_for_each_entry(counter, &evsel_list->entries, node) {
 361                         read_counter(counter);
 362                         perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
 363                 }
 364         } else {
 365                 list_for_each_entry(counter, &evsel_list->entries, node) {
 366                         read_counter_aggr(counter);
 367                         perf_evsel__close_fd(counter, evsel_list->cpus->nr,
 368                                              evsel_list->threads->nr);
 369                 }
 370         }
 371
 372         return WEXITSTATUS(status);
 373 }
 374
 375 static void print_noise(struct perf_evsel *evsel, double avg)
 376 {
 377         struct perf_stat *ps;
 378
 379         if (run_count == 1)
 380                 return;
 381
 382         ps = evsel->priv;
 383         fprintf(stderr, "   ( +- %7.3f%% )",
 384                         100 * stddev_stats(&ps->res_stats[0]) / avg);
 385 }
 386
 387 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 388 {
 389         double msecs = avg / 1e6;
 390         char cpustr[16] = { '\0', };
 391         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
 392
 393         if (no_aggr)
 394                 sprintf(cpustr, "CPU%*d%s",
 395                         csv_output ? 0 : -4,
 396                         evsel_list->cpus->map[cpu], csv_sep);
 397
 398         fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
 399
 400         if (evsel->cgrp)
 401                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 402
 403         if (csv_output)
 404                 return;
 405
 406         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 407                 fprintf(stderr, " # %10.3f CPUs ",
 408                                 avg / avg_stats(&walltime_nsecs_stats));
 409 }
 410
 411 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 412 {
 413         double total, ratio = 0.0;
 414         char cpustr[16] = { '\0', };
 415         const char *fmt;
 416
 417         if (csv_output)
 418                 fmt = "%s%.0f%s%s";
 419         else if (big_num)
 420                 fmt = "%s%'18.0f%s%-24s";
 421         else
 422                 fmt = "%s%18.0f%s%-24s";
 423
 424         if (no_aggr)
 425                 sprintf(cpustr, "CPU%*d%s",
 426                         csv_output ? 0 : -4,
 427                         evsel_list->cpus->map[cpu], csv_sep);
 428         else
 429                 cpu = 0;
 430
 431         fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
 432
 433         if (evsel->cgrp)
 434                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 435
 436         if (csv_output)
 437                 return;
 438
 439         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 440                 total = avg_stats(&runtime_cycles_stats[cpu]);
 441
 442                 if (total)
 443                         ratio = avg / total;
 444
 445                 fprintf(stderr, " # %10.3f IPC  ", ratio);
 446         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 447                         runtime_branches_stats[cpu].n != 0) {
 448                 total = avg_stats(&runtime_branches_stats[cpu]);
 449
 450                 if (total)
 451                         ratio = avg * 100 / total;
 452
 453                 fprintf(stderr, " # %10.3f %%    ", ratio);
 454
 455         } else if (runtime_nsecs_stats[cpu].n != 0) {
 456                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 457
 458                 if (total)
 459                         ratio = 1000.0 * avg / total;
 460
 461                 fprintf(stderr, " # %10.3f M/sec", ratio);
 462         }
 463 }
 464
 465 /*
 466  * Print out the results of a single counter:
 467  * aggregated counts in system-wide mode
 468  */
 469 static void print_counter_aggr(struct perf_evsel *counter)
 470 {
 471         struct perf_stat *ps = counter->priv;
 472         double avg = avg_stats(&ps->res_stats[0]);
 473         int scaled = counter->counts->scaled;
 474
 475         if (scaled == -1) {
 476                 fprintf(stderr, "%*s%s%*s",
 477                         csv_output ? 0 : 18,
 478                         "<not counted>",
 479                         csv_sep,
 480                         csv_output ? 0 : -24,
 481                         event_name(counter));
 482
 483                 if (counter->cgrp)
 484                         fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 485
 486                 fputc('\n', stderr);
 487                 return;
 488         }
 489
 490         if (nsec_counter(counter))
 491                 nsec_printout(-1, counter, avg);
 492         else
 493                 abs_printout(-1, counter, avg);
 494
 495         if (csv_output) {
 496                 fputc('\n', stderr);
 497                 return;
 498         }
 499
 500         print_noise(counter, avg);
 501
 502         if (scaled) {
 503                 double avg_enabled, avg_running;
 504
 505                 avg_enabled = avg_stats(&ps->res_stats[1]);
 506                 avg_running = avg_stats(&ps->res_stats[2]);
 507
 508                 fprintf(stderr, "  (scaled from %.2f%%)",
 509                                 100 * avg_running / avg_enabled);
 510         }
 511         fprintf(stderr, "\n");
 512 }
 513
 514 /*
 515  * Print out the results of a single counter:
 516  * does not use aggregated count in system-wide
 517  */
 518 static void print_counter(struct perf_evsel *counter)
 519 {
 520         u64 ena, run, val;
 521         int cpu;
 522
 523         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 524                 val = counter->counts->cpu[cpu].val;
 525                 ena = counter->counts->cpu[cpu].ena;
 526                 run = counter->counts->cpu[cpu].run;
 527                 if (run == 0 || ena == 0) {
 528                         fprintf(stderr, "CPU%*d%s%*s%s%*s",
 529                                 csv_output ? 0 : -4,
 530                                 evsel_list->cpus->map[cpu], csv_sep,
 531                                 csv_output ? 0 : 18,
 532                                 "<not counted>", csv_sep,
 533                                 csv_output ? 0 : -24,
 534                                 event_name(counter));
 535
 536                         if (counter->cgrp)
 537                                 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 538
 539                         fputc('\n', stderr);
 540                         continue;
 541                 }
 542
 543                 if (nsec_counter(counter))
 544                         nsec_printout(cpu, counter, val);
 545                 else
 546                         abs_printout(cpu, counter, val);
 547
 548                 if (!csv_output) {
 549                         print_noise(counter, 1.0);
 550
 551                         if (run != ena) {
 552                                 fprintf(stderr, "  (scaled from %.2f%%)",
 553                                         100.0 * run / ena);
 554                         }
 555                 }
 556                 fputc('\n', stderr);
 557         }
 558 }
 559
 560 static void print_stat(int argc, const char **argv)
 561 {
 562         struct perf_evsel *counter;
 563         int i;
 564
 565         fflush(stdout);
 566
 567         if (!csv_output) {
 568                 fprintf(stderr, "\n");
 569                 fprintf(stderr, " Performance counter stats for ");
 570                 if(target_pid == -1 && target_tid == -1) {
 571                         fprintf(stderr, "\'%s", argv[0]);
 572                         for (i = 1; i < argc; i++)
 573                                 fprintf(stderr, " %s", argv[i]);
 574                 } else if (target_pid != -1)
 575                         fprintf(stderr, "process id \'%d", target_pid);
 576                 else
 577                         fprintf(stderr, "thread id \'%d", target_tid);
 578
 579                 fprintf(stderr, "\'");
 580                 if (run_count > 1)
 581                         fprintf(stderr, " (%d runs)", run_count);
 582                 fprintf(stderr, ":\n\n");
 583         }
 584
 585         if (no_aggr) {
 586                 list_for_each_entry(counter, &evsel_list->entries, node)
 587                         print_counter(counter);
 588         } else {
 589                 list_for_each_entry(counter, &evsel_list->entries, node)
 590                         print_counter_aggr(counter);
 591         }
 592
 593         if (!csv_output) {
 594                 fprintf(stderr, "\n");
 595                 fprintf(stderr, " %18.9f  seconds time elapsed",
 596                                 avg_stats(&walltime_nsecs_stats)/1e9);
 597                 if (run_count > 1) {
 598                         fprintf(stderr, "   ( +- %7.3f%% )",
 599                                 100*stddev_stats(&walltime_nsecs_stats) /
 600                                 avg_stats(&walltime_nsecs_stats));
 601                 }
 602                 fprintf(stderr, "\n\n");
 603         }
 604 }
 605
 606 static volatile int signr = -1;
 607
 608 static void skip_signal(int signo)
 609 {
 610         if(child_pid == -1)
 611                 done = 1;
 612
 613         signr = signo;
 614 }
 615
 616 static void sig_atexit(void)
 617 {
 618         if (child_pid != -1)
 619                 kill(child_pid, SIGTERM);
 620
 621         if (signr == -1)
 622                 return;
 623
 624         signal(signr, SIG_DFL);
 625         kill(getpid(), signr);
 626 }
 627
 628 static const char * const stat_usage[] = {
 629         "perf stat [<options>] [<command>]",
 630         NULL
 631 };
 632
 633 static int stat__set_big_num(const struct option *opt __used,
 634                              const char *s __used, int unset)
 635 {
 636         big_num_opt = unset ? 0 : 1;
 637         return 0;
 638 }
 639
 640 static const struct option options[] = {
 641         OPT_CALLBACK('e', "event", &evsel_list, "event",
 642                      "event selector. use 'perf list' to list available events",
 643                      parse_events),
 644         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
 645                      "event filter", parse_filter),
 646         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 647                     "child tasks do not inherit counters"),
 648         OPT_INTEGER('p', "pid", &target_pid,
 649                     "stat events on existing process id"),
 650         OPT_INTEGER('t', "tid", &target_tid,
 651                     "stat events on existing thread id"),
 652         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 653                     "system-wide collection from all CPUs"),
 654         OPT_BOOLEAN('c', "scale", &scale,
 655                     "scale/normalize counters"),
 656         OPT_INCR('v', "verbose", &verbose,
 657                     "be more verbose (show counter open errors, etc)"),
 658         OPT_INTEGER('r', "repeat", &run_count,
 659                     "repeat command and print average + stddev (max: 100)"),
 660         OPT_BOOLEAN('n', "null", &null_run,
 661                     "null run - dont start any counters"),
 662         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
 663                            "print large numbers with thousands\' separators",
 664                            stat__set_big_num),
 665         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 666                     "list of cpus to monitor in system-wide"),
 667         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
 668                     "disable CPU count aggregation"),
 669         OPT_STRING('x', "field-separator", &csv_sep, "separator",
 670                    "print counts with custom separator"),
 671         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 672                      "monitor event in cgroup name only",
 673                      parse_cgroups),
 674         OPT_END()
 675 };
 676
 677 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 678 {
 679         struct perf_evsel *pos;
 680         int status = -ENOMEM;
 681
 682         setlocale(LC_ALL, "");
 683
 684         evsel_list = perf_evlist__new(NULL, NULL);
 685         if (evsel_list == NULL)
 686                 return -ENOMEM;
 687
 688         argc = parse_options(argc, argv, options, stat_usage,
 689                 PARSE_OPT_STOP_AT_NON_OPTION);
 690
 691         if (csv_sep)
 692                 csv_output = true;
 693         else
 694                 csv_sep = DEFAULT_SEPARATOR;
 695
 696         /*
 697          * let the spreadsheet do the pretty-printing
 698          */
 699         if (csv_output) {
 700                 /* User explicitely passed -B? */
 701                 if (big_num_opt == 1) {
 702                         fprintf(stderr, "-B option not supported with -x\n");
 703                         usage_with_options(stat_usage, options);
 704                 } else /* Nope, so disable big number formatting */
 705                         big_num = false;
 706         } else if (big_num_opt == 0) /* User passed --no-big-num */
 707                 big_num = false;
 708
 709         if (!argc && target_pid == -1 && target_tid == -1)
 710                 usage_with_options(stat_usage, options);
 711         if (run_count <= 0)
 712                 usage_with_options(stat_usage, options);
 713
 714         /* no_aggr, cgroup are for system-wide only */
 715         if ((no_aggr || nr_cgroups) && !system_wide) {
 716                 fprintf(stderr, "both cgroup and no-aggregation "
 717                         "modes only available in system-wide mode\n");
 718
 719                 usage_with_options(stat_usage, options);
 720         }
 721
 722         /* Set attrs and nr_counters if no event is selected and !null_run */
 723         if (!null_run && !evsel_list->nr_entries) {
 724                 size_t c;
 725
 726                 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
 727                         pos = perf_evsel__new(&default_attrs[c], c);
 728                         if (pos == NULL)
 729                                 goto out;
 730                         perf_evlist__add(evsel_list, pos);
 731                 }
 732         }
 733
 734         if (target_pid != -1)
 735                 target_tid = target_pid;
 736
 737         evsel_list->threads = thread_map__new(target_pid, target_tid);
 738         if (evsel_list->threads == NULL) {
 739                 pr_err("Problems finding threads of monitor\n");
 740                 usage_with_options(stat_usage, options);
 741         }
 742
 743         if (system_wide)
 744                 evsel_list->cpus = cpu_map__new(cpu_list);
 745         else
 746                 evsel_list->cpus = cpu_map__dummy_new();
 747
 748         if (evsel_list->cpus == NULL) {
 749                 perror("failed to parse CPUs map");
 750                 usage_with_options(stat_usage, options);
 751                 return -1;
 752         }
 753
 754         list_for_each_entry(pos, &evsel_list->entries, node) {
 755                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
 756                     perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
 757                     perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
 758                         goto out_free_fd;
 759         }
 760
 761         /*
 762          * We dont want to block the signals - that would cause
 763          * child tasks to inherit that and Ctrl-C would not work.
 764          * What we want is for Ctrl-C to work in the exec()-ed
 765          * task, but being ignored by perf stat itself:
 766          */
 767         atexit(sig_atexit);
 768         signal(SIGINT,  skip_signal);
 769         signal(SIGALRM, skip_signal);
 770         signal(SIGABRT, skip_signal);
 771
 772         status = 0;
 773         for (run_idx = 0; run_idx < run_count; run_idx++) {
 774                 if (run_count != 1 && verbose)
 775                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 776                 status = run_perf_stat(argc, argv);
 777         }
 778
 779         if (status != -1)
 780                 print_stat(argc, argv);
 781 out_free_fd:
 782         list_for_each_entry(pos, &evsel_list->entries, node)
 783                 perf_evsel__free_stat_priv(pos);
 784         perf_evlist__delete_maps(evsel_list);
 785 out:
 786         perf_evlist__delete(evsel_list);
 787         return status;
 788 }