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
  53 #include <sys/prctl.h>
  54 #include <math.h>
  55 #include <locale.h>
  56
  57 #define DEFAULT_SEPARATOR       " "
  58
  59 static struct perf_event_attr default_attrs[] = {
  60
  61   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
  62   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
  63   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
  64   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
  65
  66   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
  67   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
  68   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
  69   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
  70   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES        },
  71   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES            },
  72
  73 };
  74
  75 struct perf_evlist              *evsel_list;
  76
  77 static bool                     system_wide                     =  false;
  78 static struct cpu_map           *cpus;
  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 struct thread_map        *threads;
  88 static pid_t                    child_pid                       = -1;
  89 static bool                     null_run                        =  false;
  90 static bool                     big_num                         =  true;
  91 static int                      big_num_opt                     =  -1;
  92 static const char               *cpu_list;
  93 static const char               *csv_sep                        = NULL;
  94 static bool                     csv_output                      = false;
  95
  96 static volatile int done = 0;
  97
  98 struct stats
  99 {
 100         double n, mean, M2;
 101 };
 102
 103 struct perf_stat {
 104         struct stats      res_stats[3];
 105 };
 106
 107 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 108 {
 109         evsel->priv = zalloc(sizeof(struct perf_stat));
 110         return evsel->priv == NULL ? -ENOMEM : 0;
 111 }
 112
 113 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 114 {
 115         free(evsel->priv);
 116         evsel->priv = NULL;
 117 }
 118
 119 static void update_stats(struct stats *stats, u64 val)
 120 {
 121         double delta;
 122
 123         stats->n++;
 124         delta = val - stats->mean;
 125         stats->mean += delta / stats->n;
 126         stats->M2 += delta*(val - stats->mean);
 127 }
 128
 129 static double avg_stats(struct stats *stats)
 130 {
 131         return stats->mean;
 132 }
 133
 134 /*
 135  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 136  *
 137  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 138  * s^2 = -------------------------------
 139  *                  n - 1
 140  *
 141  * http://en.wikipedia.org/wiki/Stddev
 142  *
 143  * The std dev of the mean is related to the std dev by:
 144  *
 145  *             s
 146  * s_mean = -------
 147  *          sqrt(n)
 148  *
 149  */
 150 static double stddev_stats(struct stats *stats)
 151 {
 152         double variance = stats->M2 / (stats->n - 1);
 153         double variance_mean = variance / stats->n;
 154
 155         return sqrt(variance_mean);
 156 }
 157
 158 struct stats                    runtime_nsecs_stats[MAX_NR_CPUS];
 159 struct stats                    runtime_cycles_stats[MAX_NR_CPUS];
 160 struct stats                    runtime_branches_stats[MAX_NR_CPUS];
 161 struct stats                    walltime_nsecs_stats;
 162
 163 static int create_perf_stat_counter(struct perf_evsel *evsel)
 164 {
 165         struct perf_event_attr *attr = &evsel->attr;
 166
 167         if (scale)
 168                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 169                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 170
 171         if (system_wide)
 172                 return perf_evsel__open_per_cpu(evsel, cpus, false, false);
 173
 174         attr->inherit = !no_inherit;
 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, threads, false, 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, cpus->nr, 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 < 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                         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, 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, cpus->nr, threads->nr);
 361                 }
 362         }
 363
 364         return WEXITSTATUS(status);
 365 }
 366
 367 static void print_noise(struct perf_evsel *evsel, double avg)
 368 {
 369         struct perf_stat *ps;
 370
 371         if (run_count == 1)
 372                 return;
 373
 374         ps = evsel->priv;
 375         fprintf(stderr, "   ( +- %7.3f%% )",
 376                         100 * stddev_stats(&ps->res_stats[0]) / avg);
 377 }
 378
 379 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 380 {
 381         double msecs = avg / 1e6;
 382         char cpustr[16] = { '\0', };
 383         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
 384
 385         if (no_aggr)
 386                 sprintf(cpustr, "CPU%*d%s",
 387                         csv_output ? 0 : -4,
 388                         cpus->map[cpu], csv_sep);
 389
 390         fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
 391
 392         if (csv_output)
 393                 return;
 394
 395         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 396                 fprintf(stderr, " # %10.3f CPUs ",
 397                                 avg / avg_stats(&walltime_nsecs_stats));
 398 }
 399
 400 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 401 {
 402         double total, ratio = 0.0;
 403         char cpustr[16] = { '\0', };
 404         const char *fmt;
 405
 406         if (csv_output)
 407                 fmt = "%s%.0f%s%s";
 408         else if (big_num)
 409                 fmt = "%s%'18.0f%s%-24s";
 410         else
 411                 fmt = "%s%18.0f%s%-24s";
 412
 413         if (no_aggr)
 414                 sprintf(cpustr, "CPU%*d%s",
 415                         csv_output ? 0 : -4,
 416                         cpus->map[cpu], csv_sep);
 417         else
 418                 cpu = 0;
 419
 420         fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
 421
 422         if (csv_output)
 423                 return;
 424
 425         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 426                 total = avg_stats(&runtime_cycles_stats[cpu]);
 427
 428                 if (total)
 429                         ratio = avg / total;
 430
 431                 fprintf(stderr, " # %10.3f IPC  ", ratio);
 432         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 433                         runtime_branches_stats[cpu].n != 0) {
 434                 total = avg_stats(&runtime_branches_stats[cpu]);
 435
 436                 if (total)
 437                         ratio = avg * 100 / total;
 438
 439                 fprintf(stderr, " # %10.3f %%    ", ratio);
 440
 441         } else if (runtime_nsecs_stats[cpu].n != 0) {
 442                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 443
 444                 if (total)
 445                         ratio = 1000.0 * avg / total;
 446
 447                 fprintf(stderr, " # %10.3f M/sec", ratio);
 448         }
 449 }
 450
 451 /*
 452  * Print out the results of a single counter:
 453  * aggregated counts in system-wide mode
 454  */
 455 static void print_counter_aggr(struct perf_evsel *counter)
 456 {
 457         struct perf_stat *ps = counter->priv;
 458         double avg = avg_stats(&ps->res_stats[0]);
 459         int scaled = counter->counts->scaled;
 460
 461         if (scaled == -1) {
 462                 fprintf(stderr, "%*s%s%-24s\n",
 463                         csv_output ? 0 : 18,
 464                         "<not counted>", csv_sep, event_name(counter));
 465                 return;
 466         }
 467
 468         if (nsec_counter(counter))
 469                 nsec_printout(-1, counter, avg);
 470         else
 471                 abs_printout(-1, counter, avg);
 472
 473         if (csv_output) {
 474                 fputc('\n', stderr);
 475                 return;
 476         }
 477
 478         print_noise(counter, avg);
 479
 480         if (scaled) {
 481                 double avg_enabled, avg_running;
 482
 483                 avg_enabled = avg_stats(&ps->res_stats[1]);
 484                 avg_running = avg_stats(&ps->res_stats[2]);
 485
 486                 fprintf(stderr, "  (scaled from %.2f%%)",
 487                                 100 * avg_running / avg_enabled);
 488         }
 489
 490         fprintf(stderr, "\n");
 491 }
 492
 493 /*
 494  * Print out the results of a single counter:
 495  * does not use aggregated count in system-wide
 496  */
 497 static void print_counter(struct perf_evsel *counter)
 498 {
 499         u64 ena, run, val;
 500         int cpu;
 501
 502         for (cpu = 0; cpu < cpus->nr; cpu++) {
 503                 val = counter->counts->cpu[cpu].val;
 504                 ena = counter->counts->cpu[cpu].ena;
 505                 run = counter->counts->cpu[cpu].run;
 506                 if (run == 0 || ena == 0) {
 507                         fprintf(stderr, "CPU%*d%s%*s%s%-24s",
 508                                 csv_output ? 0 : -4,
 509                                 cpus->map[cpu], csv_sep,
 510                                 csv_output ? 0 : 18,
 511                                 "<not counted>", csv_sep,
 512                                 event_name(counter));
 513
 514                         fprintf(stderr, "\n");
 515                         continue;
 516                 }
 517
 518                 if (nsec_counter(counter))
 519                         nsec_printout(cpu, counter, val);
 520                 else
 521                         abs_printout(cpu, counter, val);
 522
 523                 if (!csv_output) {
 524                         print_noise(counter, 1.0);
 525
 526                         if (run != ena) {
 527                                 fprintf(stderr, "  (scaled from %.2f%%)",
 528                                         100.0 * run / ena);
 529                         }
 530                 }
 531                 fprintf(stderr, "\n");
 532         }
 533 }
 534
 535 static void print_stat(int argc, const char **argv)
 536 {
 537         struct perf_evsel *counter;
 538         int i;
 539
 540         fflush(stdout);
 541
 542         if (!csv_output) {
 543                 fprintf(stderr, "\n");
 544                 fprintf(stderr, " Performance counter stats for ");
 545                 if(target_pid == -1 && target_tid == -1) {
 546                         fprintf(stderr, "\'%s", argv[0]);
 547                         for (i = 1; i < argc; i++)
 548                                 fprintf(stderr, " %s", argv[i]);
 549                 } else if (target_pid != -1)
 550                         fprintf(stderr, "process id \'%d", target_pid);
 551                 else
 552                         fprintf(stderr, "thread id \'%d", target_tid);
 553
 554                 fprintf(stderr, "\'");
 555                 if (run_count > 1)
 556                         fprintf(stderr, " (%d runs)", run_count);
 557                 fprintf(stderr, ":\n\n");
 558         }
 559
 560         if (no_aggr) {
 561                 list_for_each_entry(counter, &evsel_list->entries, node)
 562                         print_counter(counter);
 563         } else {
 564                 list_for_each_entry(counter, &evsel_list->entries, node)
 565                         print_counter_aggr(counter);
 566         }
 567
 568         if (!csv_output) {
 569                 fprintf(stderr, "\n");
 570                 fprintf(stderr, " %18.9f  seconds time elapsed",
 571                                 avg_stats(&walltime_nsecs_stats)/1e9);
 572                 if (run_count > 1) {
 573                         fprintf(stderr, "   ( +- %7.3f%% )",
 574                                 100*stddev_stats(&walltime_nsecs_stats) /
 575                                 avg_stats(&walltime_nsecs_stats));
 576                 }
 577                 fprintf(stderr, "\n\n");
 578         }
 579 }
 580
 581 static volatile int signr = -1;
 582
 583 static void skip_signal(int signo)
 584 {
 585         if(child_pid == -1)
 586                 done = 1;
 587
 588         signr = signo;
 589 }
 590
 591 static void sig_atexit(void)
 592 {
 593         if (child_pid != -1)
 594                 kill(child_pid, SIGTERM);
 595
 596         if (signr == -1)
 597                 return;
 598
 599         signal(signr, SIG_DFL);
 600         kill(getpid(), signr);
 601 }
 602
 603 static const char * const stat_usage[] = {
 604         "perf stat [<options>] [<command>]",
 605         NULL
 606 };
 607
 608 static int stat__set_big_num(const struct option *opt __used,
 609                              const char *s __used, int unset)
 610 {
 611         big_num_opt = unset ? 0 : 1;
 612         return 0;
 613 }
 614
 615 static const struct option options[] = {
 616         OPT_CALLBACK('e', "event", &evsel_list, "event",
 617                      "event selector. use 'perf list' to list available events",
 618                      parse_events),
 619         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 620                     "child tasks do not inherit counters"),
 621         OPT_INTEGER('p', "pid", &target_pid,
 622                     "stat events on existing process id"),
 623         OPT_INTEGER('t', "tid", &target_tid,
 624                     "stat events on existing thread id"),
 625         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 626                     "system-wide collection from all CPUs"),
 627         OPT_BOOLEAN('c', "scale", &scale,
 628                     "scale/normalize counters"),
 629         OPT_INCR('v', "verbose", &verbose,
 630                     "be more verbose (show counter open errors, etc)"),
 631         OPT_INTEGER('r', "repeat", &run_count,
 632                     "repeat command and print average + stddev (max: 100)"),
 633         OPT_BOOLEAN('n', "null", &null_run,
 634                     "null run - dont start any counters"),
 635         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
 636                            "print large numbers with thousands\' separators",
 637                            stat__set_big_num),
 638         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 639                     "list of cpus to monitor in system-wide"),
 640         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
 641                     "disable CPU count aggregation"),
 642         OPT_STRING('x', "field-separator", &csv_sep, "separator",
 643                    "print counts with custom separator"),
 644         OPT_END()
 645 };
 646
 647 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 648 {
 649         struct perf_evsel *pos;
 650         int status = -ENOMEM;
 651
 652         setlocale(LC_ALL, "");
 653
 654         evsel_list = perf_evlist__new();
 655         if (evsel_list == NULL)
 656                 return -ENOMEM;
 657
 658         argc = parse_options(argc, argv, options, stat_usage,
 659                 PARSE_OPT_STOP_AT_NON_OPTION);
 660
 661         if (csv_sep)
 662                 csv_output = true;
 663         else
 664                 csv_sep = DEFAULT_SEPARATOR;
 665
 666         /*
 667          * let the spreadsheet do the pretty-printing
 668          */
 669         if (csv_output) {
 670                 /* User explicitely passed -B? */
 671                 if (big_num_opt == 1) {
 672                         fprintf(stderr, "-B option not supported with -x\n");
 673                         usage_with_options(stat_usage, options);
 674                 } else /* Nope, so disable big number formatting */
 675                         big_num = false;
 676         } else if (big_num_opt == 0) /* User passed --no-big-num */
 677                 big_num = false;
 678
 679         if (!argc && target_pid == -1 && target_tid == -1)
 680                 usage_with_options(stat_usage, options);
 681         if (run_count <= 0)
 682                 usage_with_options(stat_usage, options);
 683
 684         /* no_aggr is for system-wide only */
 685         if (no_aggr && !system_wide)
 686                 usage_with_options(stat_usage, options);
 687
 688         /* Set attrs and nr_counters if no event is selected and !null_run */
 689         if (!null_run && !evsel_list->nr_entries) {
 690                 size_t c;
 691
 692                 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
 693                         pos = perf_evsel__new(&default_attrs[c], c);
 694                         if (pos == NULL)
 695                                 goto out;
 696                         perf_evlist__add(evsel_list, pos);
 697                 }
 698         }
 699
 700         if (target_pid != -1)
 701                 target_tid = target_pid;
 702
 703         threads = thread_map__new(target_pid, target_tid);
 704         if (threads == NULL) {
 705                 pr_err("Problems finding threads of monitor\n");
 706                 usage_with_options(stat_usage, options);
 707         }
 708
 709         if (system_wide)
 710                 cpus = cpu_map__new(cpu_list);
 711         else
 712                 cpus = cpu_map__dummy_new();
 713
 714         if (cpus == NULL) {
 715                 perror("failed to parse CPUs map");
 716                 usage_with_options(stat_usage, options);
 717                 return -1;
 718         }
 719
 720         list_for_each_entry(pos, &evsel_list->entries, node) {
 721                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
 722                     perf_evsel__alloc_counts(pos, cpus->nr) < 0 ||
 723                     perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0)
 724                         goto out_free_fd;
 725         }
 726
 727         /*
 728          * We dont want to block the signals - that would cause
 729          * child tasks to inherit that and Ctrl-C would not work.
 730          * What we want is for Ctrl-C to work in the exec()-ed
 731          * task, but being ignored by perf stat itself:
 732          */
 733         atexit(sig_atexit);
 734         signal(SIGINT,  skip_signal);
 735         signal(SIGALRM, skip_signal);
 736         signal(SIGABRT, skip_signal);
 737
 738         status = 0;
 739         for (run_idx = 0; run_idx < run_count; run_idx++) {
 740                 if (run_count != 1 && verbose)
 741                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 742                 status = run_perf_stat(argc, argv);
 743         }
 744
 745         if (status != -1)
 746                 print_stat(argc, argv);
 747 out_free_fd:
 748         list_for_each_entry(pos, &evsel_list->entries, node)
 749                 perf_evsel__free_stat_priv(pos);
 750         perf_evlist__delete(evsel_list);
 751 out:
 752         thread_map__delete(threads);
 753         threads = NULL;
 754         return status;
 755 }