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