Pileus Git - ~andy/linux/blob - 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         /*
 337          * Enable counters and exec the command:
 338          */
 339         t0 = rdclock();
 340
 341         if (forks) {
 342                 close(go_pipe[1]);
 343                 wait(&status);
 344         } else {
 345                 while(!done) sleep(1);
 346         }
 347
 348         t1 = rdclock();
 349
 350         update_stats(&walltime_nsecs_stats, t1 - t0);
 351
 352         if (no_aggr) {
 353                 list_for_each_entry(counter, &evsel_list->entries, node) {
 354                         read_counter(counter);
 355                         perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
 356                 }
 357         } else {
 358                 list_for_each_entry(counter, &evsel_list->entries, node) {
 359                         read_counter_aggr(counter);
 360                         perf_evsel__close_fd(counter, evsel_list->cpus->nr,
 361                                              evsel_list->threads->nr);
 362                 }
 363         }
 364
 365         return WEXITSTATUS(status);
 366 }
 367
 368 static void print_noise(struct perf_evsel *evsel, double avg)
 369 {
 370         struct perf_stat *ps;
 371
 372         if (run_count == 1)
 373                 return;
 374
 375         ps = evsel->priv;
 376         fprintf(stderr, "   ( +- %7.3f%% )",
 377                         100 * stddev_stats(&ps->res_stats[0]) / avg);
 378 }
 379
 380 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 381 {
 382         double msecs = avg / 1e6;
 383         char cpustr[16] = { '\0', };
 384         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
 385
 386         if (no_aggr)
 387                 sprintf(cpustr, "CPU%*d%s",
 388                         csv_output ? 0 : -4,
 389                         evsel_list->cpus->map[cpu], csv_sep);
 390
 391         fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
 392
 393         if (evsel->cgrp)
 394                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 395
 396         if (csv_output)
 397                 return;
 398
 399         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 400                 fprintf(stderr, " # %10.3f CPUs ",
 401                                 avg / avg_stats(&walltime_nsecs_stats));
 402 }
 403
 404 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 405 {
 406         double total, ratio = 0.0;
 407         char cpustr[16] = { '\0', };
 408         const char *fmt;
 409
 410         if (csv_output)
 411                 fmt = "%s%.0f%s%s";
 412         else if (big_num)
 413                 fmt = "%s%'18.0f%s%-24s";
 414         else
 415                 fmt = "%s%18.0f%s%-24s";
 416
 417         if (no_aggr)
 418                 sprintf(cpustr, "CPU%*d%s",
 419                         csv_output ? 0 : -4,
 420                         evsel_list->cpus->map[cpu], csv_sep);
 421         else
 422                 cpu = 0;
 423
 424         fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
 425
 426         if (evsel->cgrp)
 427                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 428
 429         if (csv_output)
 430                 return;
 431
 432         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 433                 total = avg_stats(&runtime_cycles_stats[cpu]);
 434
 435                 if (total)
 436                         ratio = avg / total;
 437
 438                 fprintf(stderr, " # %10.3f IPC  ", ratio);
 439         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 440                         runtime_branches_stats[cpu].n != 0) {
 441                 total = avg_stats(&runtime_branches_stats[cpu]);
 442
 443                 if (total)
 444                         ratio = avg * 100 / total;
 445
 446                 fprintf(stderr, " # %10.3f %%    ", ratio);
 447
 448         } else if (runtime_nsecs_stats[cpu].n != 0) {
 449                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 450
 451                 if (total)
 452                         ratio = 1000.0 * avg / total;
 453
 454                 fprintf(stderr, " # %10.3f M/sec", ratio);
 455         }
 456 }
 457
 458 /*
 459  * Print out the results of a single counter:
 460  * aggregated counts in system-wide mode
 461  */
 462 static void print_counter_aggr(struct perf_evsel *counter)
 463 {
 464         struct perf_stat *ps = counter->priv;
 465         double avg = avg_stats(&ps->res_stats[0]);
 466         int scaled = counter->counts->scaled;
 467
 468         if (scaled == -1) {
 469                 fprintf(stderr, "%*s%s%*s",
 470                         csv_output ? 0 : 18,
 471                         "<not counted>",
 472                         csv_sep,
 473                         csv_output ? 0 : -24,
 474                         event_name(counter));
 475
 476                 if (counter->cgrp)
 477                         fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 478
 479                 fputc('\n', stderr);
 480                 return;
 481         }
 482
 483         if (nsec_counter(counter))
 484                 nsec_printout(-1, counter, avg);
 485         else
 486                 abs_printout(-1, counter, avg);
 487
 488         if (csv_output) {
 489                 fputc('\n', stderr);
 490                 return;
 491         }
 492
 493         print_noise(counter, avg);
 494
 495         if (scaled) {
 496                 double avg_enabled, avg_running;
 497
 498                 avg_enabled = avg_stats(&ps->res_stats[1]);
 499                 avg_running = avg_stats(&ps->res_stats[2]);
 500
 501                 fprintf(stderr, "  (scaled from %.2f%%)",
 502                                 100 * avg_running / avg_enabled);
 503         }
 504         fprintf(stderr, "\n");
 505 }
 506
 507 /*
 508  * Print out the results of a single counter:
 509  * does not use aggregated count in system-wide
 510  */
 511 static void print_counter(struct perf_evsel *counter)
 512 {
 513         u64 ena, run, val;
 514         int cpu;
 515
 516         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 517                 val = counter->counts->cpu[cpu].val;
 518                 ena = counter->counts->cpu[cpu].ena;
 519                 run = counter->counts->cpu[cpu].run;
 520                 if (run == 0 || ena == 0) {
 521                         fprintf(stderr, "CPU%*d%s%*s%s%*s",
 522                                 csv_output ? 0 : -4,
 523                                 evsel_list->cpus->map[cpu], csv_sep,
 524                                 csv_output ? 0 : 18,
 525                                 "<not counted>", csv_sep,
 526                                 csv_output ? 0 : -24,
 527                                 event_name(counter));
 528
 529                         if (counter->cgrp)
 530                                 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 531
 532                         fputc('\n', stderr);
 533                         continue;
 534                 }
 535
 536                 if (nsec_counter(counter))
 537                         nsec_printout(cpu, counter, val);
 538                 else
 539                         abs_printout(cpu, counter, val);
 540
 541                 if (!csv_output) {
 542                         print_noise(counter, 1.0);
 543
 544                         if (run != ena) {
 545                                 fprintf(stderr, "  (scaled from %.2f%%)",
 546                                         100.0 * run / ena);
 547                         }
 548                 }
 549                 fputc('\n', stderr);
 550         }
 551 }
 552
 553 static void print_stat(int argc, const char **argv)
 554 {
 555         struct perf_evsel *counter;
 556         int i;
 557
 558         fflush(stdout);
 559
 560         if (!csv_output) {
 561                 fprintf(stderr, "\n");
 562                 fprintf(stderr, " Performance counter stats for ");
 563                 if(target_pid == -1 && target_tid == -1) {
 564                         fprintf(stderr, "\'%s", argv[0]);
 565                         for (i = 1; i < argc; i++)
 566                                 fprintf(stderr, " %s", argv[i]);
 567                 } else if (target_pid != -1)
 568                         fprintf(stderr, "process id \'%d", target_pid);
 569                 else
 570                         fprintf(stderr, "thread id \'%d", target_tid);
 571
 572                 fprintf(stderr, "\'");
 573                 if (run_count > 1)
 574                         fprintf(stderr, " (%d runs)", run_count);
 575                 fprintf(stderr, ":\n\n");
 576         }
 577
 578         if (no_aggr) {
 579                 list_for_each_entry(counter, &evsel_list->entries, node)
 580                         print_counter(counter);
 581         } else {
 582                 list_for_each_entry(counter, &evsel_list->entries, node)
 583                         print_counter_aggr(counter);
 584         }
 585
 586         if (!csv_output) {
 587                 fprintf(stderr, "\n");
 588                 fprintf(stderr, " %18.9f  seconds time elapsed",
 589                                 avg_stats(&walltime_nsecs_stats)/1e9);
 590                 if (run_count > 1) {
 591                         fprintf(stderr, "   ( +- %7.3f%% )",
 592                                 100*stddev_stats(&walltime_nsecs_stats) /
 593                                 avg_stats(&walltime_nsecs_stats));
 594                 }
 595                 fprintf(stderr, "\n\n");
 596         }
 597 }
 598
 599 static volatile int signr = -1;
 600
 601 static void skip_signal(int signo)
 602 {
 603         if(child_pid == -1)
 604                 done = 1;
 605
 606         signr = signo;
 607 }
 608
 609 static void sig_atexit(void)
 610 {
 611         if (child_pid != -1)
 612                 kill(child_pid, SIGTERM);
 613
 614         if (signr == -1)
 615                 return;
 616
 617         signal(signr, SIG_DFL);
 618         kill(getpid(), signr);
 619 }
 620
 621 static const char * const stat_usage[] = {
 622         "perf stat [<options>] [<command>]",
 623         NULL
 624 };
 625
 626 static int stat__set_big_num(const struct option *opt __used,
 627                              const char *s __used, int unset)
 628 {
 629         big_num_opt = unset ? 0 : 1;
 630         return 0;
 631 }
 632
 633 static const struct option options[] = {
 634         OPT_CALLBACK('e', "event", &evsel_list, "event",
 635                      "event selector. use 'perf list' to list available events",
 636                      parse_events),
 637         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 638                     "child tasks do not inherit counters"),
 639         OPT_INTEGER('p', "pid", &target_pid,
 640                     "stat events on existing process id"),
 641         OPT_INTEGER('t', "tid", &target_tid,
 642                     "stat events on existing thread id"),
 643         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 644                     "system-wide collection from all CPUs"),
 645         OPT_BOOLEAN('c', "scale", &scale,
 646                     "scale/normalize counters"),
 647         OPT_INCR('v', "verbose", &verbose,
 648                     "be more verbose (show counter open errors, etc)"),
 649         OPT_INTEGER('r', "repeat", &run_count,
 650                     "repeat command and print average + stddev (max: 100)"),
 651         OPT_BOOLEAN('n', "null", &null_run,
 652                     "null run - dont start any counters"),
 653         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
 654                            "print large numbers with thousands\' separators",
 655                            stat__set_big_num),
 656         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 657                     "list of cpus to monitor in system-wide"),
 658         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
 659                     "disable CPU count aggregation"),
 660         OPT_STRING('x', "field-separator", &csv_sep, "separator",
 661                    "print counts with custom separator"),
 662         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 663                      "monitor event in cgroup name only",
 664                      parse_cgroups),
 665         OPT_END()
 666 };
 667
 668 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 669 {
 670         struct perf_evsel *pos;
 671         int status = -ENOMEM;
 672
 673         setlocale(LC_ALL, "");
 674
 675         evsel_list = perf_evlist__new(NULL, NULL);
 676         if (evsel_list == NULL)
 677                 return -ENOMEM;
 678
 679         argc = parse_options(argc, argv, options, stat_usage,
 680                 PARSE_OPT_STOP_AT_NON_OPTION);
 681
 682         if (csv_sep)
 683                 csv_output = true;
 684         else
 685                 csv_sep = DEFAULT_SEPARATOR;
 686
 687         /*
 688          * let the spreadsheet do the pretty-printing
 689          */
 690         if (csv_output) {
 691                 /* User explicitely passed -B? */
 692                 if (big_num_opt == 1) {
 693                         fprintf(stderr, "-B option not supported with -x\n");
 694                         usage_with_options(stat_usage, options);
 695                 } else /* Nope, so disable big number formatting */
 696                         big_num = false;
 697         } else if (big_num_opt == 0) /* User passed --no-big-num */
 698                 big_num = false;
 699
 700         if (!argc && target_pid == -1 && target_tid == -1)
 701                 usage_with_options(stat_usage, options);
 702         if (run_count <= 0)
 703                 usage_with_options(stat_usage, options);
 704
 705         /* no_aggr, cgroup are for system-wide only */
 706         if ((no_aggr || nr_cgroups) && !system_wide) {
 707                 fprintf(stderr, "both cgroup and no-aggregation "
 708                         "modes only available in system-wide mode\n");
 709
 710                 usage_with_options(stat_usage, options);
 711         }
 712
 713         /* Set attrs and nr_counters if no event is selected and !null_run */
 714         if (!null_run && !evsel_list->nr_entries) {
 715                 size_t c;
 716
 717                 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
 718                         pos = perf_evsel__new(&default_attrs[c], c);
 719                         if (pos == NULL)
 720                                 goto out;
 721                         perf_evlist__add(evsel_list, pos);
 722                 }
 723         }
 724
 725         if (target_pid != -1)
 726                 target_tid = target_pid;
 727
 728         evsel_list->threads = thread_map__new(target_pid, target_tid);
 729         if (evsel_list->threads == NULL) {
 730                 pr_err("Problems finding threads of monitor\n");
 731                 usage_with_options(stat_usage, options);
 732         }
 733
 734         if (system_wide)
 735                 evsel_list->cpus = cpu_map__new(cpu_list);
 736         else
 737                 evsel_list->cpus = cpu_map__dummy_new();
 738
 739         if (evsel_list->cpus == NULL) {
 740                 perror("failed to parse CPUs map");
 741                 usage_with_options(stat_usage, options);
 742                 return -1;
 743         }
 744
 745         list_for_each_entry(pos, &evsel_list->entries, node) {
 746                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
 747                     perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
 748                     perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
 749                         goto out_free_fd;
 750         }
 751
 752         /*
 753          * We dont want to block the signals - that would cause
 754          * child tasks to inherit that and Ctrl-C would not work.
 755          * What we want is for Ctrl-C to work in the exec()-ed
 756          * task, but being ignored by perf stat itself:
 757          */
 758         atexit(sig_atexit);
 759         signal(SIGINT,  skip_signal);
 760         signal(SIGALRM, skip_signal);
 761         signal(SIGABRT, skip_signal);
 762
 763         status = 0;
 764         for (run_idx = 0; run_idx < run_count; run_idx++) {
 765                 if (run_count != 1 && verbose)
 766                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 767                 status = run_perf_stat(argc, argv);
 768         }
 769
 770         if (status != -1)
 771                 print_stat(argc, argv);
 772 out_free_fd:
 773         list_for_each_entry(pos, &evsel_list->entries, node)
 774                 perf_evsel__free_stat_priv(pos);
 775         perf_evlist__delete_maps(evsel_list);
 776 out:
 777         perf_evlist__delete(evsel_list);
 778         return status;
 779 }