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
  11   Time: 0.118
  12
  13   Performance counter stats for './hackbench 10':
  14
  15        1708.761321 task-clock                #   11.037 CPUs utilized
  16             41,190 context-switches          #    0.024 M/sec
  17              6,735 CPU-migrations            #    0.004 M/sec
  18             17,318 page-faults               #    0.010 M/sec
  19      5,205,202,243 cycles                    #    3.046 GHz
  20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  22      2,603,501,247 instructions              #    0.50  insns per cycle
  23                                              #    1.48  stalled cycles per insn
  24        484,357,498 branches                  #  283.455 M/sec
  25          6,388,934 branch-misses             #    1.32% of all branches
  26
  27         0.154822978  seconds time elapsed
  28
  29  *
  30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  31  *
  32  * Improvements and fixes by:
  33  *
  34  *   Arjan van de Ven <arjan@linux.intel.com>
  35  *   Yanmin Zhang <yanmin.zhang@intel.com>
  36  *   Wu Fengguang <fengguang.wu@intel.com>
  37  *   Mike Galbraith <efault@gmx.de>
  38  *   Paul Mackerras <paulus@samba.org>
  39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  40  *
  41  * Released under the GPL v2. (and only v2, not any later version)
  42  */
  43
  44 #include "perf.h"
  45 #include "builtin.h"
  46 #include "util/util.h"
  47 #include "util/parse-options.h"
  48 #include "util/parse-events.h"
  49 #include "util/event.h"
  50 #include "util/evlist.h"
  51 #include "util/evsel.h"
  52 #include "util/debug.h"
  53 #include "util/color.h"
  54 #include "util/header.h"
  55 #include "util/cpumap.h"
  56 #include "util/thread.h"
  57 #include "util/thread_map.h"
  58
  59 #include <sys/prctl.h>
  60 #include <math.h>
  61 #include <locale.h>
  62
  63 #define DEFAULT_SEPARATOR       " "
  64 #define CNTR_NOT_SUPPORTED      "<not supported>"
  65 #define CNTR_NOT_COUNTED        "<not counted>"
  66
  67 static struct perf_event_attr default_attrs[] = {
  68
  69   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
  70   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
  71   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
  72   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
  73
  74   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
  75   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
  76   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
  77   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
  78   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
  79   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
  80
  81 };
  82
  83 /*
  84  * Detailed stats (-d), covering the L1 and last level data caches:
  85  */
  86 static struct perf_event_attr detailed_attrs[] = {
  87
  88   { .type = PERF_TYPE_HW_CACHE,
  89     .config =
  90          PERF_COUNT_HW_CACHE_L1D                <<  0  |
  91         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
  92         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
  93
  94   { .type = PERF_TYPE_HW_CACHE,
  95     .config =
  96          PERF_COUNT_HW_CACHE_L1D                <<  0  |
  97         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
  98         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
  99
 100   { .type = PERF_TYPE_HW_CACHE,
 101     .config =
 102          PERF_COUNT_HW_CACHE_LL                 <<  0  |
 103         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 104         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
 105
 106   { .type = PERF_TYPE_HW_CACHE,
 107     .config =
 108          PERF_COUNT_HW_CACHE_LL                 <<  0  |
 109         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 110         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
 111 };
 112
 113 /*
 114  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
 115  */
 116 static struct perf_event_attr very_detailed_attrs[] = {
 117
 118   { .type = PERF_TYPE_HW_CACHE,
 119     .config =
 120          PERF_COUNT_HW_CACHE_L1I                <<  0  |
 121         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 122         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
 123
 124   { .type = PERF_TYPE_HW_CACHE,
 125     .config =
 126          PERF_COUNT_HW_CACHE_L1I                <<  0  |
 127         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 128         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
 129
 130   { .type = PERF_TYPE_HW_CACHE,
 131     .config =
 132          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
 133         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 134         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
 135
 136   { .type = PERF_TYPE_HW_CACHE,
 137     .config =
 138          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
 139         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 140         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
 141
 142   { .type = PERF_TYPE_HW_CACHE,
 143     .config =
 144          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
 145         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 146         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
 147
 148   { .type = PERF_TYPE_HW_CACHE,
 149     .config =
 150          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
 151         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 152         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
 153
 154 };
 155
 156 /*
 157  * Very, very detailed stats (-d -d -d), adding prefetch events:
 158  */
 159 static struct perf_event_attr very_very_detailed_attrs[] = {
 160
 161   { .type = PERF_TYPE_HW_CACHE,
 162     .config =
 163          PERF_COUNT_HW_CACHE_L1D                <<  0  |
 164         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
 165         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
 166
 167   { .type = PERF_TYPE_HW_CACHE,
 168     .config =
 169          PERF_COUNT_HW_CACHE_L1D                <<  0  |
 170         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
 171         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
 172 };
 173
 174
 175
 176 static struct perf_evlist       *evsel_list;
 177
 178 static struct perf_target       target = {
 179         .uid    = UINT_MAX,
 180 };
 181
 182 static int                      run_idx                         =  0;
 183 static int                      run_count                       =  1;
 184 static bool                     no_inherit                      = false;
 185 static bool                     scale                           =  true;
 186 static bool                     no_aggr                         = false;
 187 static pid_t                    child_pid                       = -1;
 188 static bool                     null_run                        =  false;
 189 static int                      detailed_run                    =  0;
 190 static bool                     sync_run                        =  false;
 191 static bool                     big_num                         =  true;
 192 static int                      big_num_opt                     =  -1;
 193 static const char               *csv_sep                        = NULL;
 194 static bool                     csv_output                      = false;
 195 static bool                     group                           = false;
 196 static const char               *output_name                    = NULL;
 197 static FILE                     *output                         = NULL;
 198 static int                      output_fd;
 199
 200 static volatile int done = 0;
 201
 202 struct stats
 203 {
 204         double n, mean, M2;
 205 };
 206
 207 struct perf_stat {
 208         struct stats      res_stats[3];
 209 };
 210
 211 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 212 {
 213         evsel->priv = zalloc(sizeof(struct perf_stat));
 214         return evsel->priv == NULL ? -ENOMEM : 0;
 215 }
 216
 217 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 218 {
 219         free(evsel->priv);
 220         evsel->priv = NULL;
 221 }
 222
 223 static void update_stats(struct stats *stats, u64 val)
 224 {
 225         double delta;
 226
 227         stats->n++;
 228         delta = val - stats->mean;
 229         stats->mean += delta / stats->n;
 230         stats->M2 += delta*(val - stats->mean);
 231 }
 232
 233 static double avg_stats(struct stats *stats)
 234 {
 235         return stats->mean;
 236 }
 237
 238 /*
 239  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 240  *
 241  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 242  * s^2 = -------------------------------
 243  *                  n - 1
 244  *
 245  * http://en.wikipedia.org/wiki/Stddev
 246  *
 247  * The std dev of the mean is related to the std dev by:
 248  *
 249  *             s
 250  * s_mean = -------
 251  *          sqrt(n)
 252  *
 253  */
 254 static double stddev_stats(struct stats *stats)
 255 {
 256         double variance, variance_mean;
 257
 258         if (!stats->n)
 259                 return 0.0;
 260
 261         variance = stats->M2 / (stats->n - 1);
 262         variance_mean = variance / stats->n;
 263
 264         return sqrt(variance_mean);
 265 }
 266
 267 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
 268 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
 269 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
 270 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
 271 static struct stats runtime_branches_stats[MAX_NR_CPUS];
 272 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
 273 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
 274 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
 275 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
 276 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
 277 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
 278 static struct stats walltime_nsecs_stats;
 279
 280 static int create_perf_stat_counter(struct perf_evsel *evsel,
 281                                     struct perf_evsel *first)
 282 {
 283         struct perf_event_attr *attr = &evsel->attr;
 284         bool exclude_guest_missing = false;
 285         int ret;
 286
 287         if (scale)
 288                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 289                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 290
 291         attr->inherit = !no_inherit;
 292
 293 retry:
 294         if (exclude_guest_missing)
 295                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
 296
 297         if (perf_target__has_cpu(&target)) {
 298                 ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus);
 299                 if (ret)
 300                         goto check_ret;
 301                 return 0;
 302         }
 303
 304         if (!perf_target__has_task(&target) && (!group || evsel == first)) {
 305                 attr->disabled = 1;
 306                 attr->enable_on_exec = 1;
 307         }
 308
 309         ret = perf_evsel__open_per_thread(evsel, evsel_list->threads);
 310         if (!ret)
 311                 return 0;
 312         /* fall through */
 313 check_ret:
 314         if (ret && errno == EINVAL) {
 315                 if (!exclude_guest_missing &&
 316                     (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
 317                         pr_debug("Old kernel, cannot exclude "
 318                                  "guest or host samples.\n");
 319                         exclude_guest_missing = true;
 320                         goto retry;
 321                 }
 322         }
 323         return ret;
 324 }
 325
 326 /*
 327  * Does the counter have nsecs as a unit?
 328  */
 329 static inline int nsec_counter(struct perf_evsel *evsel)
 330 {
 331         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 332             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 333                 return 1;
 334
 335         return 0;
 336 }
 337
 338 /*
 339  * Update various tracking values we maintain to print
 340  * more semantic information such as miss/hit ratios,
 341  * instruction rates, etc:
 342  */
 343 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
 344 {
 345         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 346                 update_stats(&runtime_nsecs_stats[0], count[0]);
 347         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 348                 update_stats(&runtime_cycles_stats[0], count[0]);
 349         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
 350                 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
 351         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 352                 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
 353         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 354                 update_stats(&runtime_branches_stats[0], count[0]);
 355         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 356                 update_stats(&runtime_cacherefs_stats[0], count[0]);
 357         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
 358                 update_stats(&runtime_l1_dcache_stats[0], count[0]);
 359         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
 360                 update_stats(&runtime_l1_icache_stats[0], count[0]);
 361         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
 362                 update_stats(&runtime_ll_cache_stats[0], count[0]);
 363         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
 364                 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
 365         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
 366                 update_stats(&runtime_itlb_cache_stats[0], count[0]);
 367 }
 368
 369 /*
 370  * Read out the results of a single counter:
 371  * aggregate counts across CPUs in system-wide mode
 372  */
 373 static int read_counter_aggr(struct perf_evsel *counter)
 374 {
 375         struct perf_stat *ps = counter->priv;
 376         u64 *count = counter->counts->aggr.values;
 377         int i;
 378
 379         if (__perf_evsel__read(counter, evsel_list->cpus->nr,
 380                                evsel_list->threads->nr, scale) < 0)
 381                 return -1;
 382
 383         for (i = 0; i < 3; i++)
 384                 update_stats(&ps->res_stats[i], count[i]);
 385
 386         if (verbose) {
 387                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 388                         perf_evsel__name(counter), count[0], count[1], count[2]);
 389         }
 390
 391         /*
 392          * Save the full runtime - to allow normalization during printout:
 393          */
 394         update_shadow_stats(counter, count);
 395
 396         return 0;
 397 }
 398
 399 /*
 400  * Read out the results of a single counter:
 401  * do not aggregate counts across CPUs in system-wide mode
 402  */
 403 static int read_counter(struct perf_evsel *counter)
 404 {
 405         u64 *count;
 406         int cpu;
 407
 408         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 409                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 410                         return -1;
 411
 412                 count = counter->counts->cpu[cpu].values;
 413
 414                 update_shadow_stats(counter, count);
 415         }
 416
 417         return 0;
 418 }
 419
 420 static int run_perf_stat(int argc __maybe_unused, const char **argv)
 421 {
 422         unsigned long long t0, t1;
 423         struct perf_evsel *counter, *first;
 424         int status = 0;
 425         int child_ready_pipe[2], go_pipe[2];
 426         const bool forks = (argc > 0);
 427         char buf;
 428
 429         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 430                 perror("failed to create pipes");
 431                 return -1;
 432         }
 433
 434         if (forks) {
 435                 if ((child_pid = fork()) < 0)
 436                         perror("failed to fork");
 437
 438                 if (!child_pid) {
 439                         close(child_ready_pipe[0]);
 440                         close(go_pipe[1]);
 441                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 442
 443                         /*
 444                          * Do a dummy execvp to get the PLT entry resolved,
 445                          * so we avoid the resolver overhead on the real
 446                          * execvp call.
 447                          */
 448                         execvp("", (char **)argv);
 449
 450                         /*
 451                          * Tell the parent we're ready to go
 452                          */
 453                         close(child_ready_pipe[1]);
 454
 455                         /*
 456                          * Wait until the parent tells us to go.
 457                          */
 458                         if (read(go_pipe[0], &buf, 1) == -1)
 459                                 perror("unable to read pipe");
 460
 461                         execvp(argv[0], (char **)argv);
 462
 463                         perror(argv[0]);
 464                         exit(-1);
 465                 }
 466
 467                 if (perf_target__none(&target))
 468                         evsel_list->threads->map[0] = child_pid;
 469
 470                 /*
 471                  * Wait for the child to be ready to exec.
 472                  */
 473                 close(child_ready_pipe[1]);
 474                 close(go_pipe[0]);
 475                 if (read(child_ready_pipe[0], &buf, 1) == -1)
 476                         perror("unable to read pipe");
 477                 close(child_ready_pipe[0]);
 478         }
 479
 480         if (group)
 481                 perf_evlist__set_leader(evsel_list);
 482
 483         first = perf_evlist__first(evsel_list);
 484
 485         list_for_each_entry(counter, &evsel_list->entries, node) {
 486                 if (create_perf_stat_counter(counter, first) < 0) {
 487                         /*
 488                          * PPC returns ENXIO for HW counters until 2.6.37
 489                          * (behavior changed with commit b0a873e).
 490                          */
 491                         if (errno == EINVAL || errno == ENOSYS ||
 492                             errno == ENOENT || errno == EOPNOTSUPP ||
 493                             errno == ENXIO) {
 494                                 if (verbose)
 495                                         ui__warning("%s event is not supported by the kernel.\n",
 496                                                     perf_evsel__name(counter));
 497                                 counter->supported = false;
 498                                 continue;
 499                         }
 500
 501                         if (errno == EPERM || errno == EACCES) {
 502                                 error("You may not have permission to collect %sstats.\n"
 503                                       "\t Consider tweaking"
 504                                       " /proc/sys/kernel/perf_event_paranoid or running as root.",
 505                                       target.system_wide ? "system-wide " : "");
 506                         } else {
 507                                 error("open_counter returned with %d (%s). "
 508                                       "/bin/dmesg may provide additional information.\n",
 509                                        errno, strerror(errno));
 510                         }
 511                         if (child_pid != -1)
 512                                 kill(child_pid, SIGTERM);
 513
 514                         pr_err("Not all events could be opened.\n");
 515                         return -1;
 516                 }
 517                 counter->supported = true;
 518         }
 519
 520         if (perf_evlist__set_filters(evsel_list)) {
 521                 error("failed to set filter with %d (%s)\n", errno,
 522                         strerror(errno));
 523                 return -1;
 524         }
 525
 526         /*
 527          * Enable counters and exec the command:
 528          */
 529         t0 = rdclock();
 530
 531         if (forks) {
 532                 close(go_pipe[1]);
 533                 wait(&status);
 534                 if (WIFSIGNALED(status))
 535                         psignal(WTERMSIG(status), argv[0]);
 536         } else {
 537                 while(!done) sleep(1);
 538         }
 539
 540         t1 = rdclock();
 541
 542         update_stats(&walltime_nsecs_stats, t1 - t0);
 543
 544         if (no_aggr) {
 545                 list_for_each_entry(counter, &evsel_list->entries, node) {
 546                         read_counter(counter);
 547                         perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
 548                 }
 549         } else {
 550                 list_for_each_entry(counter, &evsel_list->entries, node) {
 551                         read_counter_aggr(counter);
 552                         perf_evsel__close_fd(counter, evsel_list->cpus->nr,
 553                                              evsel_list->threads->nr);
 554                 }
 555         }
 556
 557         return WEXITSTATUS(status);
 558 }
 559
 560 static void print_noise_pct(double total, double avg)
 561 {
 562         double pct = 0.0;
 563
 564         if (avg)
 565                 pct = 100.0*total/avg;
 566
 567         if (csv_output)
 568                 fprintf(output, "%s%.2f%%", csv_sep, pct);
 569         else if (pct)
 570                 fprintf(output, "  ( +-%6.2f%% )", pct);
 571 }
 572
 573 static void print_noise(struct perf_evsel *evsel, double avg)
 574 {
 575         struct perf_stat *ps;
 576
 577         if (run_count == 1)
 578                 return;
 579
 580         ps = evsel->priv;
 581         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 582 }
 583
 584 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 585 {
 586         double msecs = avg / 1e6;
 587         char cpustr[16] = { '\0', };
 588         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
 589
 590         if (no_aggr)
 591                 sprintf(cpustr, "CPU%*d%s",
 592                         csv_output ? 0 : -4,
 593                         evsel_list->cpus->map[cpu], csv_sep);
 594
 595         fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel));
 596
 597         if (evsel->cgrp)
 598                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 599
 600         if (csv_output)
 601                 return;
 602
 603         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 604                 fprintf(output, " # %8.3f CPUs utilized          ",
 605                         avg / avg_stats(&walltime_nsecs_stats));
 606         else
 607                 fprintf(output, "                                   ");
 608 }
 609
 610 /* used for get_ratio_color() */
 611 enum grc_type {
 612         GRC_STALLED_CYCLES_FE,
 613         GRC_STALLED_CYCLES_BE,
 614         GRC_CACHE_MISSES,
 615         GRC_MAX_NR
 616 };
 617
 618 static const char *get_ratio_color(enum grc_type type, double ratio)
 619 {
 620         static const double grc_table[GRC_MAX_NR][3] = {
 621                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
 622                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
 623                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
 624         };
 625         const char *color = PERF_COLOR_NORMAL;
 626
 627         if (ratio > grc_table[type][0])
 628                 color = PERF_COLOR_RED;
 629         else if (ratio > grc_table[type][1])
 630                 color = PERF_COLOR_MAGENTA;
 631         else if (ratio > grc_table[type][2])
 632                 color = PERF_COLOR_YELLOW;
 633
 634         return color;
 635 }
 636
 637 static void print_stalled_cycles_frontend(int cpu,
 638                                           struct perf_evsel *evsel
 639                                           __maybe_unused, double avg)
 640 {
 641         double total, ratio = 0.0;
 642         const char *color;
 643
 644         total = avg_stats(&runtime_cycles_stats[cpu]);
 645
 646         if (total)
 647                 ratio = avg / total * 100.0;
 648
 649         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
 650
 651         fprintf(output, " #  ");
 652         color_fprintf(output, color, "%6.2f%%", ratio);
 653         fprintf(output, " frontend cycles idle   ");
 654 }
 655
 656 static void print_stalled_cycles_backend(int cpu,
 657                                          struct perf_evsel *evsel
 658                                          __maybe_unused, double avg)
 659 {
 660         double total, ratio = 0.0;
 661         const char *color;
 662
 663         total = avg_stats(&runtime_cycles_stats[cpu]);
 664
 665         if (total)
 666                 ratio = avg / total * 100.0;
 667
 668         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
 669
 670         fprintf(output, " #  ");
 671         color_fprintf(output, color, "%6.2f%%", ratio);
 672         fprintf(output, " backend  cycles idle   ");
 673 }
 674
 675 static void print_branch_misses(int cpu,
 676                                 struct perf_evsel *evsel __maybe_unused,
 677                                 double avg)
 678 {
 679         double total, ratio = 0.0;
 680         const char *color;
 681
 682         total = avg_stats(&runtime_branches_stats[cpu]);
 683
 684         if (total)
 685                 ratio = avg / total * 100.0;
 686
 687         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 688
 689         fprintf(output, " #  ");
 690         color_fprintf(output, color, "%6.2f%%", ratio);
 691         fprintf(output, " of all branches        ");
 692 }
 693
 694 static void print_l1_dcache_misses(int cpu,
 695                                    struct perf_evsel *evsel __maybe_unused,
 696                                    double avg)
 697 {
 698         double total, ratio = 0.0;
 699         const char *color;
 700
 701         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
 702
 703         if (total)
 704                 ratio = avg / total * 100.0;
 705
 706         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 707
 708         fprintf(output, " #  ");
 709         color_fprintf(output, color, "%6.2f%%", ratio);
 710         fprintf(output, " of all L1-dcache hits  ");
 711 }
 712
 713 static void print_l1_icache_misses(int cpu,
 714                                    struct perf_evsel *evsel __maybe_unused,
 715                                    double avg)
 716 {
 717         double total, ratio = 0.0;
 718         const char *color;
 719
 720         total = avg_stats(&runtime_l1_icache_stats[cpu]);
 721
 722         if (total)
 723                 ratio = avg / total * 100.0;
 724
 725         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 726
 727         fprintf(output, " #  ");
 728         color_fprintf(output, color, "%6.2f%%", ratio);
 729         fprintf(output, " of all L1-icache hits  ");
 730 }
 731
 732 static void print_dtlb_cache_misses(int cpu,
 733                                     struct perf_evsel *evsel __maybe_unused,
 734                                     double avg)
 735 {
 736         double total, ratio = 0.0;
 737         const char *color;
 738
 739         total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
 740
 741         if (total)
 742                 ratio = avg / total * 100.0;
 743
 744         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 745
 746         fprintf(output, " #  ");
 747         color_fprintf(output, color, "%6.2f%%", ratio);
 748         fprintf(output, " of all dTLB cache hits ");
 749 }
 750
 751 static void print_itlb_cache_misses(int cpu,
 752                                     struct perf_evsel *evsel __maybe_unused,
 753                                     double avg)
 754 {
 755         double total, ratio = 0.0;
 756         const char *color;
 757
 758         total = avg_stats(&runtime_itlb_cache_stats[cpu]);
 759
 760         if (total)
 761                 ratio = avg / total * 100.0;
 762
 763         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 764
 765         fprintf(output, " #  ");
 766         color_fprintf(output, color, "%6.2f%%", ratio);
 767         fprintf(output, " of all iTLB cache hits ");
 768 }
 769
 770 static void print_ll_cache_misses(int cpu,
 771                                   struct perf_evsel *evsel __maybe_unused,
 772                                   double avg)
 773 {
 774         double total, ratio = 0.0;
 775         const char *color;
 776
 777         total = avg_stats(&runtime_ll_cache_stats[cpu]);
 778
 779         if (total)
 780                 ratio = avg / total * 100.0;
 781
 782         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 783
 784         fprintf(output, " #  ");
 785         color_fprintf(output, color, "%6.2f%%", ratio);
 786         fprintf(output, " of all LL-cache hits   ");
 787 }
 788
 789 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 790 {
 791         double total, ratio = 0.0;
 792         char cpustr[16] = { '\0', };
 793         const char *fmt;
 794
 795         if (csv_output)
 796                 fmt = "%s%.0f%s%s";
 797         else if (big_num)
 798                 fmt = "%s%'18.0f%s%-25s";
 799         else
 800                 fmt = "%s%18.0f%s%-25s";
 801
 802         if (no_aggr)
 803                 sprintf(cpustr, "CPU%*d%s",
 804                         csv_output ? 0 : -4,
 805                         evsel_list->cpus->map[cpu], csv_sep);
 806         else
 807                 cpu = 0;
 808
 809         fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel));
 810
 811         if (evsel->cgrp)
 812                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 813
 814         if (csv_output)
 815                 return;
 816
 817         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 818                 total = avg_stats(&runtime_cycles_stats[cpu]);
 819
 820                 if (total)
 821                         ratio = avg / total;
 822
 823                 fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
 824
 825                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
 826                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
 827
 828                 if (total && avg) {
 829                         ratio = total / avg;
 830                         fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
 831                 }
 832
 833         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 834                         runtime_branches_stats[cpu].n != 0) {
 835                 print_branch_misses(cpu, evsel, avg);
 836         } else if (
 837                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 838                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
 839                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 840                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 841                         runtime_l1_dcache_stats[cpu].n != 0) {
 842                 print_l1_dcache_misses(cpu, evsel, avg);
 843         } else if (
 844                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 845                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
 846                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 847                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 848                         runtime_l1_icache_stats[cpu].n != 0) {
 849                 print_l1_icache_misses(cpu, evsel, avg);
 850         } else if (
 851                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 852                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
 853                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 854                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 855                         runtime_dtlb_cache_stats[cpu].n != 0) {
 856                 print_dtlb_cache_misses(cpu, evsel, avg);
 857         } else if (
 858                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 859                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
 860                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 861                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 862                         runtime_itlb_cache_stats[cpu].n != 0) {
 863                 print_itlb_cache_misses(cpu, evsel, avg);
 864         } else if (
 865                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 866                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
 867                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 868                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 869                         runtime_ll_cache_stats[cpu].n != 0) {
 870                 print_ll_cache_misses(cpu, evsel, avg);
 871         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
 872                         runtime_cacherefs_stats[cpu].n != 0) {
 873                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
 874
 875                 if (total)
 876                         ratio = avg * 100 / total;
 877
 878                 fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
 879
 880         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
 881                 print_stalled_cycles_frontend(cpu, evsel, avg);
 882         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
 883                 print_stalled_cycles_backend(cpu, evsel, avg);
 884         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
 885                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 886
 887                 if (total)
 888                         ratio = 1.0 * avg / total;
 889
 890                 fprintf(output, " # %8.3f GHz                    ", ratio);
 891         } else if (runtime_nsecs_stats[cpu].n != 0) {
 892                 char unit = 'M';
 893
 894                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 895
 896                 if (total)
 897                         ratio = 1000.0 * avg / total;
 898                 if (ratio < 0.001) {
 899                         ratio *= 1000;
 900                         unit = 'K';
 901                 }
 902
 903                 fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
 904         } else {
 905                 fprintf(output, "                                   ");
 906         }
 907 }
 908
 909 /*
 910  * Print out the results of a single counter:
 911  * aggregated counts in system-wide mode
 912  */
 913 static void print_counter_aggr(struct perf_evsel *counter)
 914 {
 915         struct perf_stat *ps = counter->priv;
 916         double avg = avg_stats(&ps->res_stats[0]);
 917         int scaled = counter->counts->scaled;
 918
 919         if (scaled == -1) {
 920                 fprintf(output, "%*s%s%*s",
 921                         csv_output ? 0 : 18,
 922                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 923                         csv_sep,
 924                         csv_output ? 0 : -24,
 925                         perf_evsel__name(counter));
 926
 927                 if (counter->cgrp)
 928                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
 929
 930                 fputc('\n', output);
 931                 return;
 932         }
 933
 934         if (nsec_counter(counter))
 935                 nsec_printout(-1, counter, avg);
 936         else
 937                 abs_printout(-1, counter, avg);
 938
 939         print_noise(counter, avg);
 940
 941         if (csv_output) {
 942                 fputc('\n', output);
 943                 return;
 944         }
 945
 946         if (scaled) {
 947                 double avg_enabled, avg_running;
 948
 949                 avg_enabled = avg_stats(&ps->res_stats[1]);
 950                 avg_running = avg_stats(&ps->res_stats[2]);
 951
 952                 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
 953         }
 954         fprintf(output, "\n");
 955 }
 956
 957 /*
 958  * Print out the results of a single counter:
 959  * does not use aggregated count in system-wide
 960  */
 961 static void print_counter(struct perf_evsel *counter)
 962 {
 963         u64 ena, run, val;
 964         int cpu;
 965
 966         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 967                 val = counter->counts->cpu[cpu].val;
 968                 ena = counter->counts->cpu[cpu].ena;
 969                 run = counter->counts->cpu[cpu].run;
 970                 if (run == 0 || ena == 0) {
 971                         fprintf(output, "CPU%*d%s%*s%s%*s",
 972                                 csv_output ? 0 : -4,
 973                                 evsel_list->cpus->map[cpu], csv_sep,
 974                                 csv_output ? 0 : 18,
 975                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 976                                 csv_sep,
 977                                 csv_output ? 0 : -24,
 978                                 perf_evsel__name(counter));
 979
 980                         if (counter->cgrp)
 981                                 fprintf(output, "%s%s",
 982                                         csv_sep, counter->cgrp->name);
 983
 984                         fputc('\n', output);
 985                         continue;
 986                 }
 987
 988                 if (nsec_counter(counter))
 989                         nsec_printout(cpu, counter, val);
 990                 else
 991                         abs_printout(cpu, counter, val);
 992
 993                 if (!csv_output) {
 994                         print_noise(counter, 1.0);
 995
 996                         if (run != ena)
 997                                 fprintf(output, "  (%.2f%%)",
 998                                         100.0 * run / ena);
 999                 }
1000                 fputc('\n', output);
1001         }
1002 }
1003
1004 static void print_stat(int argc, const char **argv)
1005 {
1006         struct perf_evsel *counter;
1007         int i;
1008
1009         fflush(stdout);
1010
1011         if (!csv_output) {
1012                 fprintf(output, "\n");
1013                 fprintf(output, " Performance counter stats for ");
1014                 if (!perf_target__has_task(&target)) {
1015                         fprintf(output, "\'%s", argv[0]);
1016                         for (i = 1; i < argc; i++)
1017                                 fprintf(output, " %s", argv[i]);
1018                 } else if (target.pid)
1019                         fprintf(output, "process id \'%s", target.pid);
1020                 else
1021                         fprintf(output, "thread id \'%s", target.tid);
1022
1023                 fprintf(output, "\'");
1024                 if (run_count > 1)
1025                         fprintf(output, " (%d runs)", run_count);
1026                 fprintf(output, ":\n\n");
1027         }
1028
1029         if (no_aggr) {
1030                 list_for_each_entry(counter, &evsel_list->entries, node)
1031                         print_counter(counter);
1032         } else {
1033                 list_for_each_entry(counter, &evsel_list->entries, node)
1034                         print_counter_aggr(counter);
1035         }
1036
1037         if (!csv_output) {
1038                 if (!null_run)
1039                         fprintf(output, "\n");
1040                 fprintf(output, " %17.9f seconds time elapsed",
1041                                 avg_stats(&walltime_nsecs_stats)/1e9);
1042                 if (run_count > 1) {
1043                         fprintf(output, "                                        ");
1044                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1045                                         avg_stats(&walltime_nsecs_stats));
1046                 }
1047                 fprintf(output, "\n\n");
1048         }
1049 }
1050
1051 static volatile int signr = -1;
1052
1053 static void skip_signal(int signo)
1054 {
1055         if(child_pid == -1)
1056                 done = 1;
1057
1058         signr = signo;
1059 }
1060
1061 static void sig_atexit(void)
1062 {
1063         if (child_pid != -1)
1064                 kill(child_pid, SIGTERM);
1065
1066         if (signr == -1)
1067                 return;
1068
1069         signal(signr, SIG_DFL);
1070         kill(getpid(), signr);
1071 }
1072
1073 static const char * const stat_usage[] = {
1074         "perf stat [<options>] [<command>]",
1075         NULL
1076 };
1077
1078 static int stat__set_big_num(const struct option *opt __maybe_unused,
1079                              const char *s __maybe_unused, int unset)
1080 {
1081         big_num_opt = unset ? 0 : 1;
1082         return 0;
1083 }
1084
1085 static bool append_file;
1086
1087 static const struct option options[] = {
1088         OPT_CALLBACK('e', "event", &evsel_list, "event",
1089                      "event selector. use 'perf list' to list available events",
1090                      parse_events_option),
1091         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1092                      "event filter", parse_filter),
1093         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1094                     "child tasks do not inherit counters"),
1095         OPT_STRING('p', "pid", &target.pid, "pid",
1096                    "stat events on existing process id"),
1097         OPT_STRING('t', "tid", &target.tid, "tid",
1098                    "stat events on existing thread id"),
1099         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1100                     "system-wide collection from all CPUs"),
1101         OPT_BOOLEAN('g', "group", &group,
1102                     "put the counters into a counter group"),
1103         OPT_BOOLEAN('c', "scale", &scale,
1104                     "scale/normalize counters"),
1105         OPT_INCR('v', "verbose", &verbose,
1106                     "be more verbose (show counter open errors, etc)"),
1107         OPT_INTEGER('r', "repeat", &run_count,
1108                     "repeat command and print average + stddev (max: 100)"),
1109         OPT_BOOLEAN('n', "null", &null_run,
1110                     "null run - dont start any counters"),
1111         OPT_INCR('d', "detailed", &detailed_run,
1112                     "detailed run - start a lot of events"),
1113         OPT_BOOLEAN('S', "sync", &sync_run,
1114                     "call sync() before starting a run"),
1115         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1116                            "print large numbers with thousands\' separators",
1117                            stat__set_big_num),
1118         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1119                     "list of cpus to monitor in system-wide"),
1120         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1121                     "disable CPU count aggregation"),
1122         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1123                    "print counts with custom separator"),
1124         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1125                      "monitor event in cgroup name only",
1126                      parse_cgroups),
1127         OPT_STRING('o', "output", &output_name, "file",
1128                     "output file name"),
1129         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1130         OPT_INTEGER(0, "log-fd", &output_fd,
1131                     "log output to fd, instead of stderr"),
1132         OPT_END()
1133 };
1134
1135 /*
1136  * Add default attributes, if there were no attributes specified or
1137  * if -d/--detailed, -d -d or -d -d -d is used:
1138  */
1139 static int add_default_attributes(void)
1140 {
1141         /* Set attrs if no event is selected and !null_run: */
1142         if (null_run)
1143                 return 0;
1144
1145         if (!evsel_list->nr_entries) {
1146                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1147                         return -1;
1148         }
1149
1150         /* Detailed events get appended to the event list: */
1151
1152         if (detailed_run <  1)
1153                 return 0;
1154
1155         /* Append detailed run extra attributes: */
1156         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1157                 return -1;
1158
1159         if (detailed_run < 2)
1160                 return 0;
1161
1162         /* Append very detailed run extra attributes: */
1163         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1164                 return -1;
1165
1166         if (detailed_run < 3)
1167                 return 0;
1168
1169         /* Append very, very detailed run extra attributes: */
1170         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1171 }
1172
1173 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1174 {
1175         struct perf_evsel *pos;
1176         int status = -ENOMEM;
1177         const char *mode;
1178
1179         setlocale(LC_ALL, "");
1180
1181         evsel_list = perf_evlist__new(NULL, NULL);
1182         if (evsel_list == NULL)
1183                 return -ENOMEM;
1184
1185         argc = parse_options(argc, argv, options, stat_usage,
1186                 PARSE_OPT_STOP_AT_NON_OPTION);
1187
1188         output = stderr;
1189         if (output_name && strcmp(output_name, "-"))
1190                 output = NULL;
1191
1192         if (output_name && output_fd) {
1193                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1194                 usage_with_options(stat_usage, options);
1195         }
1196
1197         if (output_fd < 0) {
1198                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1199                 usage_with_options(stat_usage, options);
1200         }
1201
1202         if (!output) {
1203                 struct timespec tm;
1204                 mode = append_file ? "a" : "w";
1205
1206                 output = fopen(output_name, mode);
1207                 if (!output) {
1208                         perror("failed to create output file");
1209                         return -1;
1210                 }
1211                 clock_gettime(CLOCK_REALTIME, &tm);
1212                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1213         } else if (output_fd > 0) {
1214                 mode = append_file ? "a" : "w";
1215                 output = fdopen(output_fd, mode);
1216                 if (!output) {
1217                         perror("Failed opening logfd");
1218                         return -errno;
1219                 }
1220         }
1221
1222         if (csv_sep) {
1223                 csv_output = true;
1224                 if (!strcmp(csv_sep, "\\t"))
1225                         csv_sep = "\t";
1226         } else
1227                 csv_sep = DEFAULT_SEPARATOR;
1228
1229         /*
1230          * let the spreadsheet do the pretty-printing
1231          */
1232         if (csv_output) {
1233                 /* User explicitly passed -B? */
1234                 if (big_num_opt == 1) {
1235                         fprintf(stderr, "-B option not supported with -x\n");
1236                         usage_with_options(stat_usage, options);
1237                 } else /* Nope, so disable big number formatting */
1238                         big_num = false;
1239         } else if (big_num_opt == 0) /* User passed --no-big-num */
1240                 big_num = false;
1241
1242         if (!argc && !perf_target__has_task(&target))
1243                 usage_with_options(stat_usage, options);
1244         if (run_count <= 0)
1245                 usage_with_options(stat_usage, options);
1246
1247         /* no_aggr, cgroup are for system-wide only */
1248         if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) {
1249                 fprintf(stderr, "both cgroup and no-aggregation "
1250                         "modes only available in system-wide mode\n");
1251
1252                 usage_with_options(stat_usage, options);
1253         }
1254
1255         if (add_default_attributes())
1256                 goto out;
1257
1258         perf_target__validate(&target);
1259
1260         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1261                 if (perf_target__has_task(&target))
1262                         pr_err("Problems finding threads of monitor\n");
1263                 if (perf_target__has_cpu(&target))
1264                         perror("failed to parse CPUs map");
1265
1266                 usage_with_options(stat_usage, options);
1267                 return -1;
1268         }
1269
1270         list_for_each_entry(pos, &evsel_list->entries, node) {
1271                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1272                     perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0)
1273                         goto out_free_fd;
1274         }
1275
1276         /*
1277          * We dont want to block the signals - that would cause
1278          * child tasks to inherit that and Ctrl-C would not work.
1279          * What we want is for Ctrl-C to work in the exec()-ed
1280          * task, but being ignored by perf stat itself:
1281          */
1282         atexit(sig_atexit);
1283         signal(SIGINT,  skip_signal);
1284         signal(SIGALRM, skip_signal);
1285         signal(SIGABRT, skip_signal);
1286
1287         status = 0;
1288         for (run_idx = 0; run_idx < run_count; run_idx++) {
1289                 if (run_count != 1 && verbose)
1290                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1291                                 run_idx + 1);
1292
1293                 if (sync_run)
1294                         sync();
1295
1296                 status = run_perf_stat(argc, argv);
1297         }
1298
1299         if (status != -1)
1300                 print_stat(argc, argv);
1301 out_free_fd:
1302         list_for_each_entry(pos, &evsel_list->entries, node)
1303                 perf_evsel__free_stat_priv(pos);
1304         perf_evlist__delete_maps(evsel_list);
1305 out:
1306         perf_evlist__delete(evsel_list);
1307         return status;
1308 }