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/color.h"
  50 #include "util/header.h"
  51 #include "util/cpumap.h"
  52 #include "util/thread.h"
  53 #include "util/thread_map.h"
  54
  55 #include <sys/prctl.h>
  56 #include <math.h>
  57 #include <locale.h>
  58
  59 #define DEFAULT_SEPARATOR       " "
  60
  61 static struct perf_event_attr default_attrs[] = {
  62
  63   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
  64   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
  65   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
  66   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
  67
  68   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
  69   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES          },
  70   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
  71   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
  72   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_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_stalled_cycles_stats[MAX_NR_CPUS];
 160 struct stats                    runtime_branches_stats[MAX_NR_CPUS];
 161 struct stats                    runtime_cacherefs_stats[MAX_NR_CPUS];
 162 struct stats                    walltime_nsecs_stats;
 163
 164 static int create_perf_stat_counter(struct perf_evsel *evsel)
 165 {
 166         struct perf_event_attr *attr = &evsel->attr;
 167
 168         if (scale)
 169                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 170                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 171
 172         attr->inherit = !no_inherit;
 173
 174         if (system_wide)
 175                 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false);
 176
 177         if (target_pid == -1 && target_tid == -1) {
 178                 attr->disabled = 1;
 179                 attr->enable_on_exec = 1;
 180         }
 181
 182         return perf_evsel__open_per_thread(evsel, evsel_list->threads, false);
 183 }
 184
 185 /*
 186  * Does the counter have nsecs as a unit?
 187  */
 188 static inline int nsec_counter(struct perf_evsel *evsel)
 189 {
 190         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 191             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 192                 return 1;
 193
 194         return 0;
 195 }
 196
 197 /*
 198  * Update various tracking values we maintain to print
 199  * more semantic information such as miss/hit ratios,
 200  * instruction rates, etc:
 201  */
 202 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
 203 {
 204         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 205                 update_stats(&runtime_nsecs_stats[0], count[0]);
 206         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 207                 update_stats(&runtime_cycles_stats[0], count[0]);
 208         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES))
 209                 update_stats(&runtime_stalled_cycles_stats[0], count[0]);
 210         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 211                 update_stats(&runtime_branches_stats[0], count[0]);
 212         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 213                 update_stats(&runtime_cacherefs_stats[0], count[0]);
 214 }
 215
 216 /*
 217  * Read out the results of a single counter:
 218  * aggregate counts across CPUs in system-wide mode
 219  */
 220 static int read_counter_aggr(struct perf_evsel *counter)
 221 {
 222         struct perf_stat *ps = counter->priv;
 223         u64 *count = counter->counts->aggr.values;
 224         int i;
 225
 226         if (__perf_evsel__read(counter, evsel_list->cpus->nr,
 227                                evsel_list->threads->nr, scale) < 0)
 228                 return -1;
 229
 230         for (i = 0; i < 3; i++)
 231                 update_stats(&ps->res_stats[i], count[i]);
 232
 233         if (verbose) {
 234                 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 235                         event_name(counter), count[0], count[1], count[2]);
 236         }
 237
 238         /*
 239          * Save the full runtime - to allow normalization during printout:
 240          */
 241         update_shadow_stats(counter, count);
 242
 243         return 0;
 244 }
 245
 246 /*
 247  * Read out the results of a single counter:
 248  * do not aggregate counts across CPUs in system-wide mode
 249  */
 250 static int read_counter(struct perf_evsel *counter)
 251 {
 252         u64 *count;
 253         int cpu;
 254
 255         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 256                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 257                         return -1;
 258
 259                 count = counter->counts->cpu[cpu].values;
 260
 261                 update_shadow_stats(counter, count);
 262         }
 263
 264         return 0;
 265 }
 266
 267 static int run_perf_stat(int argc __used, const char **argv)
 268 {
 269         unsigned long long t0, t1;
 270         struct perf_evsel *counter;
 271         int status = 0;
 272         int child_ready_pipe[2], go_pipe[2];
 273         const bool forks = (argc > 0);
 274         char buf;
 275
 276         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 277                 perror("failed to create pipes");
 278                 exit(1);
 279         }
 280
 281         if (forks) {
 282                 if ((child_pid = fork()) < 0)
 283                         perror("failed to fork");
 284
 285                 if (!child_pid) {
 286                         close(child_ready_pipe[0]);
 287                         close(go_pipe[1]);
 288                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 289
 290                         /*
 291                          * Do a dummy execvp to get the PLT entry resolved,
 292                          * so we avoid the resolver overhead on the real
 293                          * execvp call.
 294                          */
 295                         execvp("", (char **)argv);
 296
 297                         /*
 298                          * Tell the parent we're ready to go
 299                          */
 300                         close(child_ready_pipe[1]);
 301
 302                         /*
 303                          * Wait until the parent tells us to go.
 304                          */
 305                         if (read(go_pipe[0], &buf, 1) == -1)
 306                                 perror("unable to read pipe");
 307
 308                         execvp(argv[0], (char **)argv);
 309
 310                         perror(argv[0]);
 311                         exit(-1);
 312                 }
 313
 314                 if (target_tid == -1 && target_pid == -1 && !system_wide)
 315                         evsel_list->threads->map[0] = child_pid;
 316
 317                 /*
 318                  * Wait for the child to be ready to exec.
 319                  */
 320                 close(child_ready_pipe[1]);
 321                 close(go_pipe[0]);
 322                 if (read(child_ready_pipe[0], &buf, 1) == -1)
 323                         perror("unable to read pipe");
 324                 close(child_ready_pipe[0]);
 325         }
 326
 327         list_for_each_entry(counter, &evsel_list->entries, node) {
 328                 if (create_perf_stat_counter(counter) < 0) {
 329                         if (errno == -EPERM || errno == -EACCES) {
 330                                 error("You may not have permission to collect %sstats.\n"
 331                                       "\t Consider tweaking"
 332                                       " /proc/sys/kernel/perf_event_paranoid or running as root.",
 333                                       system_wide ? "system-wide " : "");
 334                         } else if (errno == ENOENT) {
 335                                 error("%s event is not supported. ", event_name(counter));
 336                         } else {
 337                                 error("open_counter returned with %d (%s). "
 338                                       "/bin/dmesg may provide additional information.\n",
 339                                        errno, strerror(errno));
 340                         }
 341                         if (child_pid != -1)
 342                                 kill(child_pid, SIGTERM);
 343                         die("Not all events could be opened.\n");
 344                         return -1;
 345                 }
 346         }
 347
 348         if (perf_evlist__set_filters(evsel_list)) {
 349                 error("failed to set filter with %d (%s)\n", errno,
 350                         strerror(errno));
 351                 return -1;
 352         }
 353
 354         /*
 355          * Enable counters and exec the command:
 356          */
 357         t0 = rdclock();
 358
 359         if (forks) {
 360                 close(go_pipe[1]);
 361                 wait(&status);
 362         } else {
 363                 while(!done) sleep(1);
 364         }
 365
 366         t1 = rdclock();
 367
 368         update_stats(&walltime_nsecs_stats, t1 - t0);
 369
 370         if (no_aggr) {
 371                 list_for_each_entry(counter, &evsel_list->entries, node) {
 372                         read_counter(counter);
 373                         perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
 374                 }
 375         } else {
 376                 list_for_each_entry(counter, &evsel_list->entries, node) {
 377                         read_counter_aggr(counter);
 378                         perf_evsel__close_fd(counter, evsel_list->cpus->nr,
 379                                              evsel_list->threads->nr);
 380                 }
 381         }
 382
 383         return WEXITSTATUS(status);
 384 }
 385
 386 static void print_noise_pct(double total, double avg)
 387 {
 388         double pct = 0.0;
 389
 390         if (avg)
 391                 pct = 100.0*total/avg;
 392
 393         fprintf(stderr, "  ( +-%6.2f%% )", pct);
 394 }
 395
 396 static void print_noise(struct perf_evsel *evsel, double avg)
 397 {
 398         struct perf_stat *ps;
 399
 400         if (run_count == 1)
 401                 return;
 402
 403         ps = evsel->priv;
 404         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 405 }
 406
 407 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 408 {
 409         double msecs = avg / 1e6;
 410         char cpustr[16] = { '\0', };
 411         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
 412
 413         if (no_aggr)
 414                 sprintf(cpustr, "CPU%*d%s",
 415                         csv_output ? 0 : -4,
 416                         evsel_list->cpus->map[cpu], csv_sep);
 417
 418         fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
 419
 420         if (evsel->cgrp)
 421                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 422
 423         if (csv_output)
 424                 return;
 425
 426         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 427                 fprintf(stderr, " # %8.3f CPUs utilized          ", avg / avg_stats(&walltime_nsecs_stats));
 428 }
 429
 430 static void print_stalled_cycles(int cpu, struct perf_evsel *evsel __used, double avg)
 431 {
 432         double total, ratio = 0.0;
 433         const char *color;
 434
 435         total = avg_stats(&runtime_cycles_stats[cpu]);
 436
 437         if (total)
 438                 ratio = avg / total * 100.0;
 439
 440         color = PERF_COLOR_NORMAL;
 441         if (ratio > 75.0)
 442                 color = PERF_COLOR_RED;
 443         else if (ratio > 50.0)
 444                 color = PERF_COLOR_MAGENTA;
 445         else if (ratio > 25.0)
 446                 color = PERF_COLOR_YELLOW;
 447
 448         fprintf(stderr, " #   ");
 449         color_fprintf(stderr, color, "%5.2f%%", ratio);
 450         fprintf(stderr, " of all cycles are idle ");
 451 }
 452
 453 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 454 {
 455         double total, ratio = 0.0;
 456         const char *color;
 457
 458         total = avg_stats(&runtime_branches_stats[cpu]);
 459
 460         if (total)
 461                 ratio = avg / total * 100.0;
 462
 463         color = PERF_COLOR_NORMAL;
 464         if (ratio > 20.0)
 465                 color = PERF_COLOR_RED;
 466         else if (ratio > 10.0)
 467                 color = PERF_COLOR_MAGENTA;
 468         else if (ratio > 5.0)
 469                 color = PERF_COLOR_YELLOW;
 470
 471         fprintf(stderr, " #   ");
 472         color_fprintf(stderr, color, "%5.2f%%", ratio);
 473         fprintf(stderr, " of all branches        ");
 474 }
 475
 476 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 477 {
 478         double total, ratio = 0.0;
 479         char cpustr[16] = { '\0', };
 480         const char *fmt;
 481
 482         if (csv_output)
 483                 fmt = "%s%.0f%s%s";
 484         else if (big_num)
 485                 fmt = "%s%'18.0f%s%-24s";
 486         else
 487                 fmt = "%s%18.0f%s%-24s";
 488
 489         if (no_aggr)
 490                 sprintf(cpustr, "CPU%*d%s",
 491                         csv_output ? 0 : -4,
 492                         evsel_list->cpus->map[cpu], csv_sep);
 493         else
 494                 cpu = 0;
 495
 496         fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
 497
 498         if (evsel->cgrp)
 499                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 500
 501         if (csv_output)
 502                 return;
 503
 504         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 505                 total = avg_stats(&runtime_cycles_stats[cpu]);
 506
 507                 if (total)
 508                         ratio = avg / total;
 509
 510                 fprintf(stderr, " #    %4.2f  insns per cycle", ratio);
 511
 512                 total = avg_stats(&runtime_stalled_cycles_stats[cpu]);
 513
 514                 if (total && avg) {
 515                         ratio = total / avg;
 516                         fprintf(stderr, "\n                                            #    %4.2f  stalled cycles per insn", ratio);
 517                 }
 518
 519         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 520                         runtime_branches_stats[cpu].n != 0) {
 521                 print_branch_misses(cpu, evsel, avg);
 522         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
 523                         runtime_cacherefs_stats[cpu].n != 0) {
 524                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
 525
 526                 if (total)
 527                         ratio = avg * 100 / total;
 528
 529                 fprintf(stderr, " # %8.3f %% of all cache refs    ", ratio);
 530
 531         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES)) {
 532                 print_stalled_cycles(cpu, evsel, avg);
 533         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
 534                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 535
 536                 if (total)
 537                         ratio = 1.0 * avg / total;
 538
 539                 fprintf(stderr, " # %8.3f GHz                    ", ratio);
 540         } else if (runtime_nsecs_stats[cpu].n != 0) {
 541                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 542
 543                 if (total)
 544                         ratio = 1000.0 * avg / total;
 545
 546                 fprintf(stderr, " # %8.3f M/sec                  ", ratio);
 547         } else {
 548                 fprintf(stderr, "                                   ");
 549         }
 550 }
 551
 552 /*
 553  * Print out the results of a single counter:
 554  * aggregated counts in system-wide mode
 555  */
 556 static void print_counter_aggr(struct perf_evsel *counter)
 557 {
 558         struct perf_stat *ps = counter->priv;
 559         double avg = avg_stats(&ps->res_stats[0]);
 560         int scaled = counter->counts->scaled;
 561
 562         if (scaled == -1) {
 563                 fprintf(stderr, "%*s%s%*s",
 564                         csv_output ? 0 : 18,
 565                         "<not counted>",
 566                         csv_sep,
 567                         csv_output ? 0 : -24,
 568                         event_name(counter));
 569
 570                 if (counter->cgrp)
 571                         fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 572
 573                 fputc('\n', stderr);
 574                 return;
 575         }
 576
 577         if (nsec_counter(counter))
 578                 nsec_printout(-1, counter, avg);
 579         else
 580                 abs_printout(-1, counter, avg);
 581
 582         if (csv_output) {
 583                 fputc('\n', stderr);
 584                 return;
 585         }
 586
 587         print_noise(counter, avg);
 588
 589         if (scaled) {
 590                 double avg_enabled, avg_running;
 591
 592                 avg_enabled = avg_stats(&ps->res_stats[1]);
 593                 avg_running = avg_stats(&ps->res_stats[2]);
 594
 595                 fprintf(stderr, "  (scaled from %.2f%%)",
 596                                 100 * avg_running / avg_enabled);
 597         }
 598         fprintf(stderr, "\n");
 599 }
 600
 601 /*
 602  * Print out the results of a single counter:
 603  * does not use aggregated count in system-wide
 604  */
 605 static void print_counter(struct perf_evsel *counter)
 606 {
 607         u64 ena, run, val;
 608         int cpu;
 609
 610         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 611                 val = counter->counts->cpu[cpu].val;
 612                 ena = counter->counts->cpu[cpu].ena;
 613                 run = counter->counts->cpu[cpu].run;
 614                 if (run == 0 || ena == 0) {
 615                         fprintf(stderr, "CPU%*d%s%*s%s%*s",
 616                                 csv_output ? 0 : -4,
 617                                 evsel_list->cpus->map[cpu], csv_sep,
 618                                 csv_output ? 0 : 18,
 619                                 "<not counted>", csv_sep,
 620                                 csv_output ? 0 : -24,
 621                                 event_name(counter));
 622
 623                         if (counter->cgrp)
 624                                 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 625
 626                         fputc('\n', stderr);
 627                         continue;
 628                 }
 629
 630                 if (nsec_counter(counter))
 631                         nsec_printout(cpu, counter, val);
 632                 else
 633                         abs_printout(cpu, counter, val);
 634
 635                 if (!csv_output) {
 636                         print_noise(counter, 1.0);
 637
 638                         if (run != ena) {
 639                                 fprintf(stderr, "  (scaled from %.2f%%)",
 640                                         100.0 * run / ena);
 641                         }
 642                 }
 643                 fputc('\n', stderr);
 644         }
 645 }
 646
 647 static void print_stat(int argc, const char **argv)
 648 {
 649         struct perf_evsel *counter;
 650         int i;
 651
 652         fflush(stdout);
 653
 654         if (!csv_output) {
 655                 fprintf(stderr, "\n");
 656                 fprintf(stderr, " Performance counter stats for ");
 657                 if(target_pid == -1 && target_tid == -1) {
 658                         fprintf(stderr, "\'%s", argv[0]);
 659                         for (i = 1; i < argc; i++)
 660                                 fprintf(stderr, " %s", argv[i]);
 661                 } else if (target_pid != -1)
 662                         fprintf(stderr, "process id \'%d", target_pid);
 663                 else
 664                         fprintf(stderr, "thread id \'%d", target_tid);
 665
 666                 fprintf(stderr, "\'");
 667                 if (run_count > 1)
 668                         fprintf(stderr, " (%d runs)", run_count);
 669                 fprintf(stderr, ":\n\n");
 670         }
 671
 672         if (no_aggr) {
 673                 list_for_each_entry(counter, &evsel_list->entries, node)
 674                         print_counter(counter);
 675         } else {
 676                 list_for_each_entry(counter, &evsel_list->entries, node)
 677                         print_counter_aggr(counter);
 678         }
 679
 680         if (!csv_output) {
 681                 fprintf(stderr, "\n");
 682                 fprintf(stderr, " %18.9f  seconds time elapsed",
 683                                 avg_stats(&walltime_nsecs_stats)/1e9);
 684                 if (run_count > 1) {
 685                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
 686                                         avg_stats(&walltime_nsecs_stats));
 687                 }
 688                 fprintf(stderr, "\n\n");
 689         }
 690 }
 691
 692 static volatile int signr = -1;
 693
 694 static void skip_signal(int signo)
 695 {
 696         if(child_pid == -1)
 697                 done = 1;
 698
 699         signr = signo;
 700 }
 701
 702 static void sig_atexit(void)
 703 {
 704         if (child_pid != -1)
 705                 kill(child_pid, SIGTERM);
 706
 707         if (signr == -1)
 708                 return;
 709
 710         signal(signr, SIG_DFL);
 711         kill(getpid(), signr);
 712 }
 713
 714 static const char * const stat_usage[] = {
 715         "perf stat [<options>] [<command>]",
 716         NULL
 717 };
 718
 719 static int stat__set_big_num(const struct option *opt __used,
 720                              const char *s __used, int unset)
 721 {
 722         big_num_opt = unset ? 0 : 1;
 723         return 0;
 724 }
 725
 726 static const struct option options[] = {
 727         OPT_CALLBACK('e', "event", &evsel_list, "event",
 728                      "event selector. use 'perf list' to list available events",
 729                      parse_events),
 730         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
 731                      "event filter", parse_filter),
 732         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 733                     "child tasks do not inherit counters"),
 734         OPT_INTEGER('p', "pid", &target_pid,
 735                     "stat events on existing process id"),
 736         OPT_INTEGER('t', "tid", &target_tid,
 737                     "stat events on existing thread id"),
 738         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 739                     "system-wide collection from all CPUs"),
 740         OPT_BOOLEAN('c', "scale", &scale,
 741                     "scale/normalize counters"),
 742         OPT_INCR('v', "verbose", &verbose,
 743                     "be more verbose (show counter open errors, etc)"),
 744         OPT_INTEGER('r', "repeat", &run_count,
 745                     "repeat command and print average + stddev (max: 100)"),
 746         OPT_BOOLEAN('n', "null", &null_run,
 747                     "null run - dont start any counters"),
 748         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
 749                            "print large numbers with thousands\' separators",
 750                            stat__set_big_num),
 751         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 752                     "list of cpus to monitor in system-wide"),
 753         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
 754                     "disable CPU count aggregation"),
 755         OPT_STRING('x', "field-separator", &csv_sep, "separator",
 756                    "print counts with custom separator"),
 757         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 758                      "monitor event in cgroup name only",
 759                      parse_cgroups),
 760         OPT_END()
 761 };
 762
 763 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 764 {
 765         struct perf_evsel *pos;
 766         int status = -ENOMEM;
 767
 768         setlocale(LC_ALL, "");
 769
 770         evsel_list = perf_evlist__new(NULL, NULL);
 771         if (evsel_list == NULL)
 772                 return -ENOMEM;
 773
 774         argc = parse_options(argc, argv, options, stat_usage,
 775                 PARSE_OPT_STOP_AT_NON_OPTION);
 776
 777         if (csv_sep)
 778                 csv_output = true;
 779         else
 780                 csv_sep = DEFAULT_SEPARATOR;
 781
 782         /*
 783          * let the spreadsheet do the pretty-printing
 784          */
 785         if (csv_output) {
 786                 /* User explicitely passed -B? */
 787                 if (big_num_opt == 1) {
 788                         fprintf(stderr, "-B option not supported with -x\n");
 789                         usage_with_options(stat_usage, options);
 790                 } else /* Nope, so disable big number formatting */
 791                         big_num = false;
 792         } else if (big_num_opt == 0) /* User passed --no-big-num */
 793                 big_num = false;
 794
 795         if (!argc && target_pid == -1 && target_tid == -1)
 796                 usage_with_options(stat_usage, options);
 797         if (run_count <= 0)
 798                 usage_with_options(stat_usage, options);
 799
 800         /* no_aggr, cgroup are for system-wide only */
 801         if ((no_aggr || nr_cgroups) && !system_wide) {
 802                 fprintf(stderr, "both cgroup and no-aggregation "
 803                         "modes only available in system-wide mode\n");
 804
 805                 usage_with_options(stat_usage, options);
 806         }
 807
 808         /* Set attrs and nr_counters if no event is selected and !null_run */
 809         if (!null_run && !evsel_list->nr_entries) {
 810                 size_t c;
 811
 812                 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
 813                         pos = perf_evsel__new(&default_attrs[c], c);
 814                         if (pos == NULL)
 815                                 goto out;
 816                         perf_evlist__add(evsel_list, pos);
 817                 }
 818         }
 819
 820         if (target_pid != -1)
 821                 target_tid = target_pid;
 822
 823         evsel_list->threads = thread_map__new(target_pid, target_tid);
 824         if (evsel_list->threads == NULL) {
 825                 pr_err("Problems finding threads of monitor\n");
 826                 usage_with_options(stat_usage, options);
 827         }
 828
 829         if (system_wide)
 830                 evsel_list->cpus = cpu_map__new(cpu_list);
 831         else
 832                 evsel_list->cpus = cpu_map__dummy_new();
 833
 834         if (evsel_list->cpus == NULL) {
 835                 perror("failed to parse CPUs map");
 836                 usage_with_options(stat_usage, options);
 837                 return -1;
 838         }
 839
 840         list_for_each_entry(pos, &evsel_list->entries, node) {
 841                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
 842                     perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
 843                     perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
 844                         goto out_free_fd;
 845         }
 846
 847         /*
 848          * We dont want to block the signals - that would cause
 849          * child tasks to inherit that and Ctrl-C would not work.
 850          * What we want is for Ctrl-C to work in the exec()-ed
 851          * task, but being ignored by perf stat itself:
 852          */
 853         atexit(sig_atexit);
 854         signal(SIGINT,  skip_signal);
 855         signal(SIGALRM, skip_signal);
 856         signal(SIGABRT, skip_signal);
 857
 858         status = 0;
 859         for (run_idx = 0; run_idx < run_count; run_idx++) {
 860                 if (run_count != 1 && verbose)
 861                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 862                 status = run_perf_stat(argc, argv);
 863         }
 864
 865         if (status != -1)
 866                 print_stat(argc, argv);
 867 out_free_fd:
 868         list_for_each_entry(pos, &evsel_list->entries, node)
 869                 perf_evsel__free_stat_priv(pos);
 870         perf_evlist__delete_maps(evsel_list);
 871 out:
 872         perf_evlist__delete(evsel_list);
 873         return status;
 874 }