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, variance_mean;
 258
 259         if (!stats->n)
 260                 return 0.0;
 261
 262         variance = stats->M2 / (stats->n - 1);
 263         variance_mean = variance / stats->n;
 264
 265         return sqrt(variance_mean);
 266 }
 267
 268 struct stats                    runtime_nsecs_stats[MAX_NR_CPUS];
 269 struct stats                    runtime_cycles_stats[MAX_NR_CPUS];
 270 struct stats                    runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
 271 struct stats                    runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
 272 struct stats                    runtime_branches_stats[MAX_NR_CPUS];
 273 struct stats                    runtime_cacherefs_stats[MAX_NR_CPUS];
 274 struct stats                    runtime_l1_dcache_stats[MAX_NR_CPUS];
 275 struct stats                    runtime_l1_icache_stats[MAX_NR_CPUS];
 276 struct stats                    runtime_ll_cache_stats[MAX_NR_CPUS];
 277 struct stats                    runtime_itlb_cache_stats[MAX_NR_CPUS];
 278 struct stats                    runtime_dtlb_cache_stats[MAX_NR_CPUS];
 279 struct stats                    walltime_nsecs_stats;
 280
 281 static int create_perf_stat_counter(struct perf_evsel *evsel)
 282 {
 283         struct perf_event_attr *attr = &evsel->attr;
 284
 285         if (scale)
 286                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 287                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 288
 289         attr->inherit = !no_inherit;
 290
 291         if (system_wide)
 292                 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, group);
 293
 294         if (target_pid == -1 && target_tid == -1) {
 295                 attr->disabled = 1;
 296                 attr->enable_on_exec = 1;
 297         }
 298
 299         return perf_evsel__open_per_thread(evsel, evsel_list->threads, group);
 300 }
 301
 302 /*
 303  * Does the counter have nsecs as a unit?
 304  */
 305 static inline int nsec_counter(struct perf_evsel *evsel)
 306 {
 307         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 308             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 309                 return 1;
 310
 311         return 0;
 312 }
 313
 314 /*
 315  * Update various tracking values we maintain to print
 316  * more semantic information such as miss/hit ratios,
 317  * instruction rates, etc:
 318  */
 319 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
 320 {
 321         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 322                 update_stats(&runtime_nsecs_stats[0], count[0]);
 323         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 324                 update_stats(&runtime_cycles_stats[0], count[0]);
 325         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
 326                 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
 327         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 328                 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
 329         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 330                 update_stats(&runtime_branches_stats[0], count[0]);
 331         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 332                 update_stats(&runtime_cacherefs_stats[0], count[0]);
 333         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
 334                 update_stats(&runtime_l1_dcache_stats[0], count[0]);
 335         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
 336                 update_stats(&runtime_l1_icache_stats[0], count[0]);
 337         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
 338                 update_stats(&runtime_ll_cache_stats[0], count[0]);
 339         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
 340                 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
 341         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
 342                 update_stats(&runtime_itlb_cache_stats[0], count[0]);
 343 }
 344
 345 /*
 346  * Read out the results of a single counter:
 347  * aggregate counts across CPUs in system-wide mode
 348  */
 349 static int read_counter_aggr(struct perf_evsel *counter)
 350 {
 351         struct perf_stat *ps = counter->priv;
 352         u64 *count = counter->counts->aggr.values;
 353         int i;
 354
 355         if (__perf_evsel__read(counter, evsel_list->cpus->nr,
 356                                evsel_list->threads->nr, scale) < 0)
 357                 return -1;
 358
 359         for (i = 0; i < 3; i++)
 360                 update_stats(&ps->res_stats[i], count[i]);
 361
 362         if (verbose) {
 363                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 364                         event_name(counter), count[0], count[1], count[2]);
 365         }
 366
 367         /*
 368          * Save the full runtime - to allow normalization during printout:
 369          */
 370         update_shadow_stats(counter, count);
 371
 372         return 0;
 373 }
 374
 375 /*
 376  * Read out the results of a single counter:
 377  * do not aggregate counts across CPUs in system-wide mode
 378  */
 379 static int read_counter(struct perf_evsel *counter)
 380 {
 381         u64 *count;
 382         int cpu;
 383
 384         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 385                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 386                         return -1;
 387
 388                 count = counter->counts->cpu[cpu].values;
 389
 390                 update_shadow_stats(counter, count);
 391         }
 392
 393         return 0;
 394 }
 395
 396 static int run_perf_stat(int argc __used, const char **argv)
 397 {
 398         unsigned long long t0, t1;
 399         struct perf_evsel *counter;
 400         int status = 0;
 401         int child_ready_pipe[2], go_pipe[2];
 402         const bool forks = (argc > 0);
 403         char buf;
 404
 405         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 406                 perror("failed to create pipes");
 407                 exit(1);
 408         }
 409
 410         if (forks) {
 411                 if ((child_pid = fork()) < 0)
 412                         perror("failed to fork");
 413
 414                 if (!child_pid) {
 415                         close(child_ready_pipe[0]);
 416                         close(go_pipe[1]);
 417                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 418
 419                         /*
 420                          * Do a dummy execvp to get the PLT entry resolved,
 421                          * so we avoid the resolver overhead on the real
 422                          * execvp call.
 423                          */
 424                         execvp("", (char **)argv);
 425
 426                         /*
 427                          * Tell the parent we're ready to go
 428                          */
 429                         close(child_ready_pipe[1]);
 430
 431                         /*
 432                          * Wait until the parent tells us to go.
 433                          */
 434                         if (read(go_pipe[0], &buf, 1) == -1)
 435                                 perror("unable to read pipe");
 436
 437                         execvp(argv[0], (char **)argv);
 438
 439                         perror(argv[0]);
 440                         exit(-1);
 441                 }
 442
 443                 if (target_tid == -1 && target_pid == -1 && !system_wide)
 444                         evsel_list->threads->map[0] = child_pid;
 445
 446                 /*
 447                  * Wait for the child to be ready to exec.
 448                  */
 449                 close(child_ready_pipe[1]);
 450                 close(go_pipe[0]);
 451                 if (read(child_ready_pipe[0], &buf, 1) == -1)
 452                         perror("unable to read pipe");
 453                 close(child_ready_pipe[0]);
 454         }
 455
 456         list_for_each_entry(counter, &evsel_list->entries, node) {
 457                 if (create_perf_stat_counter(counter) < 0) {
 458                         if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
 459                                 if (verbose)
 460                                         ui__warning("%s event is not supported by the kernel.\n",
 461                                                     event_name(counter));
 462                                 counter->supported = false;
 463                                 continue;
 464                         }
 465
 466                         if (errno == EPERM || errno == EACCES) {
 467                                 error("You may not have permission to collect %sstats.\n"
 468                                       "\t Consider tweaking"
 469                                       " /proc/sys/kernel/perf_event_paranoid or running as root.",
 470                                       system_wide ? "system-wide " : "");
 471                         } else {
 472                                 error("open_counter returned with %d (%s). "
 473                                       "/bin/dmesg may provide additional information.\n",
 474                                        errno, strerror(errno));
 475                         }
 476                         if (child_pid != -1)
 477                                 kill(child_pid, SIGTERM);
 478                         die("Not all events could be opened.\n");
 479                         return -1;
 480                 }
 481                 counter->supported = true;
 482         }
 483
 484         if (perf_evlist__set_filters(evsel_list)) {
 485                 error("failed to set filter with %d (%s)\n", errno,
 486                         strerror(errno));
 487                 return -1;
 488         }
 489
 490         /*
 491          * Enable counters and exec the command:
 492          */
 493         t0 = rdclock();
 494
 495         if (forks) {
 496                 close(go_pipe[1]);
 497                 wait(&status);
 498         } else {
 499                 while(!done) sleep(1);
 500         }
 501
 502         t1 = rdclock();
 503
 504         update_stats(&walltime_nsecs_stats, t1 - t0);
 505
 506         if (no_aggr) {
 507                 list_for_each_entry(counter, &evsel_list->entries, node) {
 508                         read_counter(counter);
 509                         perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
 510                 }
 511         } else {
 512                 list_for_each_entry(counter, &evsel_list->entries, node) {
 513                         read_counter_aggr(counter);
 514                         perf_evsel__close_fd(counter, evsel_list->cpus->nr,
 515                                              evsel_list->threads->nr);
 516                 }
 517         }
 518
 519         return WEXITSTATUS(status);
 520 }
 521
 522 static void print_noise_pct(double total, double avg)
 523 {
 524         double pct = 0.0;
 525
 526         if (avg)
 527                 pct = 100.0*total/avg;
 528
 529         if (csv_output)
 530                 fprintf(output, "%s%.2f%%", csv_sep, pct);
 531         else if (pct)
 532                 fprintf(output, "  ( +-%6.2f%% )", pct);
 533 }
 534
 535 static void print_noise(struct perf_evsel *evsel, double avg)
 536 {
 537         struct perf_stat *ps;
 538
 539         if (run_count == 1)
 540                 return;
 541
 542         ps = evsel->priv;
 543         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 544 }
 545
 546 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 547 {
 548         double msecs = avg / 1e6;
 549         char cpustr[16] = { '\0', };
 550         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
 551
 552         if (no_aggr)
 553                 sprintf(cpustr, "CPU%*d%s",
 554                         csv_output ? 0 : -4,
 555                         evsel_list->cpus->map[cpu], csv_sep);
 556
 557         fprintf(output, fmt, cpustr, msecs, csv_sep, event_name(evsel));
 558
 559         if (evsel->cgrp)
 560                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 561
 562         if (csv_output)
 563                 return;
 564
 565         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 566                 fprintf(output, " # %8.3f CPUs utilized          ",
 567                         avg / avg_stats(&walltime_nsecs_stats));
 568 }
 569
 570 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
 571 {
 572         double total, ratio = 0.0;
 573         const char *color;
 574
 575         total = avg_stats(&runtime_cycles_stats[cpu]);
 576
 577         if (total)
 578                 ratio = avg / total * 100.0;
 579
 580         color = PERF_COLOR_NORMAL;
 581         if (ratio > 50.0)
 582                 color = PERF_COLOR_RED;
 583         else if (ratio > 30.0)
 584                 color = PERF_COLOR_MAGENTA;
 585         else if (ratio > 10.0)
 586                 color = PERF_COLOR_YELLOW;
 587
 588         fprintf(output, " #  ");
 589         color_fprintf(output, color, "%6.2f%%", ratio);
 590         fprintf(output, " frontend cycles idle   ");
 591 }
 592
 593 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
 594 {
 595         double total, ratio = 0.0;
 596         const char *color;
 597
 598         total = avg_stats(&runtime_cycles_stats[cpu]);
 599
 600         if (total)
 601                 ratio = avg / total * 100.0;
 602
 603         color = PERF_COLOR_NORMAL;
 604         if (ratio > 75.0)
 605                 color = PERF_COLOR_RED;
 606         else if (ratio > 50.0)
 607                 color = PERF_COLOR_MAGENTA;
 608         else if (ratio > 20.0)
 609                 color = PERF_COLOR_YELLOW;
 610
 611         fprintf(output, " #  ");
 612         color_fprintf(output, color, "%6.2f%%", ratio);
 613         fprintf(output, " backend  cycles idle   ");
 614 }
 615
 616 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 617 {
 618         double total, ratio = 0.0;
 619         const char *color;
 620
 621         total = avg_stats(&runtime_branches_stats[cpu]);
 622
 623         if (total)
 624                 ratio = avg / total * 100.0;
 625
 626         color = PERF_COLOR_NORMAL;
 627         if (ratio > 20.0)
 628                 color = PERF_COLOR_RED;
 629         else if (ratio > 10.0)
 630                 color = PERF_COLOR_MAGENTA;
 631         else if (ratio > 5.0)
 632                 color = PERF_COLOR_YELLOW;
 633
 634         fprintf(output, " #  ");
 635         color_fprintf(output, color, "%6.2f%%", ratio);
 636         fprintf(output, " of all branches        ");
 637 }
 638
 639 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 640 {
 641         double total, ratio = 0.0;
 642         const char *color;
 643
 644         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
 645
 646         if (total)
 647                 ratio = avg / total * 100.0;
 648
 649         color = PERF_COLOR_NORMAL;
 650         if (ratio > 20.0)
 651                 color = PERF_COLOR_RED;
 652         else if (ratio > 10.0)
 653                 color = PERF_COLOR_MAGENTA;
 654         else if (ratio > 5.0)
 655                 color = PERF_COLOR_YELLOW;
 656
 657         fprintf(output, " #  ");
 658         color_fprintf(output, color, "%6.2f%%", ratio);
 659         fprintf(output, " of all L1-dcache hits  ");
 660 }
 661
 662 static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 663 {
 664         double total, ratio = 0.0;
 665         const char *color;
 666
 667         total = avg_stats(&runtime_l1_icache_stats[cpu]);
 668
 669         if (total)
 670                 ratio = avg / total * 100.0;
 671
 672         color = PERF_COLOR_NORMAL;
 673         if (ratio > 20.0)
 674                 color = PERF_COLOR_RED;
 675         else if (ratio > 10.0)
 676                 color = PERF_COLOR_MAGENTA;
 677         else if (ratio > 5.0)
 678                 color = PERF_COLOR_YELLOW;
 679
 680         fprintf(output, " #  ");
 681         color_fprintf(output, color, "%6.2f%%", ratio);
 682         fprintf(output, " of all L1-icache hits  ");
 683 }
 684
 685 static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 686 {
 687         double total, ratio = 0.0;
 688         const char *color;
 689
 690         total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
 691
 692         if (total)
 693                 ratio = avg / total * 100.0;
 694
 695         color = PERF_COLOR_NORMAL;
 696         if (ratio > 20.0)
 697                 color = PERF_COLOR_RED;
 698         else if (ratio > 10.0)
 699                 color = PERF_COLOR_MAGENTA;
 700         else if (ratio > 5.0)
 701                 color = PERF_COLOR_YELLOW;
 702
 703         fprintf(output, " #  ");
 704         color_fprintf(output, color, "%6.2f%%", ratio);
 705         fprintf(output, " of all dTLB cache hits ");
 706 }
 707
 708 static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 709 {
 710         double total, ratio = 0.0;
 711         const char *color;
 712
 713         total = avg_stats(&runtime_itlb_cache_stats[cpu]);
 714
 715         if (total)
 716                 ratio = avg / total * 100.0;
 717
 718         color = PERF_COLOR_NORMAL;
 719         if (ratio > 20.0)
 720                 color = PERF_COLOR_RED;
 721         else if (ratio > 10.0)
 722                 color = PERF_COLOR_MAGENTA;
 723         else if (ratio > 5.0)
 724                 color = PERF_COLOR_YELLOW;
 725
 726         fprintf(output, " #  ");
 727         color_fprintf(output, color, "%6.2f%%", ratio);
 728         fprintf(output, " of all iTLB cache hits ");
 729 }
 730
 731 static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 732 {
 733         double total, ratio = 0.0;
 734         const char *color;
 735
 736         total = avg_stats(&runtime_ll_cache_stats[cpu]);
 737
 738         if (total)
 739                 ratio = avg / total * 100.0;
 740
 741         color = PERF_COLOR_NORMAL;
 742         if (ratio > 20.0)
 743                 color = PERF_COLOR_RED;
 744         else if (ratio > 10.0)
 745                 color = PERF_COLOR_MAGENTA;
 746         else if (ratio > 5.0)
 747                 color = PERF_COLOR_YELLOW;
 748
 749         fprintf(output, " #  ");
 750         color_fprintf(output, color, "%6.2f%%", ratio);
 751         fprintf(output, " of all LL-cache hits   ");
 752 }
 753
 754 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 755 {
 756         double total, ratio = 0.0;
 757         char cpustr[16] = { '\0', };
 758         const char *fmt;
 759
 760         if (csv_output)
 761                 fmt = "%s%.0f%s%s";
 762         else if (big_num)
 763                 fmt = "%s%'18.0f%s%-25s";
 764         else
 765                 fmt = "%s%18.0f%s%-25s";
 766
 767         if (no_aggr)
 768                 sprintf(cpustr, "CPU%*d%s",
 769                         csv_output ? 0 : -4,
 770                         evsel_list->cpus->map[cpu], csv_sep);
 771         else
 772                 cpu = 0;
 773
 774         fprintf(output, fmt, cpustr, avg, csv_sep, event_name(evsel));
 775
 776         if (evsel->cgrp)
 777                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 778
 779         if (csv_output)
 780                 return;
 781
 782         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 783                 total = avg_stats(&runtime_cycles_stats[cpu]);
 784
 785                 if (total)
 786                         ratio = avg / total;
 787
 788                 fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
 789
 790                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
 791                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
 792
 793                 if (total && avg) {
 794                         ratio = total / avg;
 795                         fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
 796                 }
 797
 798         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 799                         runtime_branches_stats[cpu].n != 0) {
 800                 print_branch_misses(cpu, evsel, avg);
 801         } else if (
 802                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 803                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
 804                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 805                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 806                         runtime_l1_dcache_stats[cpu].n != 0) {
 807                 print_l1_dcache_misses(cpu, evsel, avg);
 808         } else if (
 809                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 810                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
 811                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 812                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 813                         runtime_l1_icache_stats[cpu].n != 0) {
 814                 print_l1_icache_misses(cpu, evsel, avg);
 815         } else if (
 816                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 817                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
 818                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 819                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 820                         runtime_dtlb_cache_stats[cpu].n != 0) {
 821                 print_dtlb_cache_misses(cpu, evsel, avg);
 822         } else if (
 823                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 824                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
 825                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 826                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 827                         runtime_itlb_cache_stats[cpu].n != 0) {
 828                 print_itlb_cache_misses(cpu, evsel, avg);
 829         } else if (
 830                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 831                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
 832                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 833                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 834                         runtime_ll_cache_stats[cpu].n != 0) {
 835                 print_ll_cache_misses(cpu, evsel, avg);
 836         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
 837                         runtime_cacherefs_stats[cpu].n != 0) {
 838                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
 839
 840                 if (total)
 841                         ratio = avg * 100 / total;
 842
 843                 fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
 844
 845         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
 846                 print_stalled_cycles_frontend(cpu, evsel, avg);
 847         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
 848                 print_stalled_cycles_backend(cpu, evsel, avg);
 849         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
 850                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 851
 852                 if (total)
 853                         ratio = 1.0 * avg / total;
 854
 855                 fprintf(output, " # %8.3f GHz                    ", ratio);
 856         } else if (runtime_nsecs_stats[cpu].n != 0) {
 857                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 858
 859                 if (total)
 860                         ratio = 1000.0 * avg / total;
 861
 862                 fprintf(output, " # %8.3f M/sec                  ", ratio);
 863         } else {
 864                 fprintf(output, "                                   ");
 865         }
 866 }
 867
 868 /*
 869  * Print out the results of a single counter:
 870  * aggregated counts in system-wide mode
 871  */
 872 static void print_counter_aggr(struct perf_evsel *counter)
 873 {
 874         struct perf_stat *ps = counter->priv;
 875         double avg = avg_stats(&ps->res_stats[0]);
 876         int scaled = counter->counts->scaled;
 877
 878         if (scaled == -1) {
 879                 fprintf(output, "%*s%s%*s",
 880                         csv_output ? 0 : 18,
 881                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 882                         csv_sep,
 883                         csv_output ? 0 : -24,
 884                         event_name(counter));
 885
 886                 if (counter->cgrp)
 887                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
 888
 889                 fputc('\n', output);
 890                 return;
 891         }
 892
 893         if (nsec_counter(counter))
 894                 nsec_printout(-1, counter, avg);
 895         else
 896                 abs_printout(-1, counter, avg);
 897
 898         print_noise(counter, avg);
 899
 900         if (csv_output) {
 901                 fputc('\n', output);
 902                 return;
 903         }
 904
 905         if (scaled) {
 906                 double avg_enabled, avg_running;
 907
 908                 avg_enabled = avg_stats(&ps->res_stats[1]);
 909                 avg_running = avg_stats(&ps->res_stats[2]);
 910
 911                 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
 912         }
 913         fprintf(output, "\n");
 914 }
 915
 916 /*
 917  * Print out the results of a single counter:
 918  * does not use aggregated count in system-wide
 919  */
 920 static void print_counter(struct perf_evsel *counter)
 921 {
 922         u64 ena, run, val;
 923         int cpu;
 924
 925         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 926                 val = counter->counts->cpu[cpu].val;
 927                 ena = counter->counts->cpu[cpu].ena;
 928                 run = counter->counts->cpu[cpu].run;
 929                 if (run == 0 || ena == 0) {
 930                         fprintf(output, "CPU%*d%s%*s%s%*s",
 931                                 csv_output ? 0 : -4,
 932                                 evsel_list->cpus->map[cpu], csv_sep,
 933                                 csv_output ? 0 : 18,
 934                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 935                                 csv_sep,
 936                                 csv_output ? 0 : -24,
 937                                 event_name(counter));
 938
 939                         if (counter->cgrp)
 940                                 fprintf(output, "%s%s",
 941                                         csv_sep, counter->cgrp->name);
 942
 943                         fputc('\n', output);
 944                         continue;
 945                 }
 946
 947                 if (nsec_counter(counter))
 948                         nsec_printout(cpu, counter, val);
 949                 else
 950                         abs_printout(cpu, counter, val);
 951
 952                 if (!csv_output) {
 953                         print_noise(counter, 1.0);
 954
 955                         if (run != ena)
 956                                 fprintf(output, "  (%.2f%%)",
 957                                         100.0 * run / ena);
 958                 }
 959                 fputc('\n', output);
 960         }
 961 }
 962
 963 static void print_stat(int argc, const char **argv)
 964 {
 965         struct perf_evsel *counter;
 966         int i;
 967
 968         fflush(stdout);
 969
 970         if (!csv_output) {
 971                 fprintf(output, "\n");
 972                 fprintf(output, " Performance counter stats for ");
 973                 if(target_pid == -1 && target_tid == -1) {
 974                         fprintf(output, "\'%s", argv[0]);
 975                         for (i = 1; i < argc; i++)
 976                                 fprintf(output, " %s", argv[i]);
 977                 } else if (target_pid != -1)
 978                         fprintf(output, "process id \'%d", target_pid);
 979                 else
 980                         fprintf(output, "thread id \'%d", target_tid);
 981
 982                 fprintf(output, "\'");
 983                 if (run_count > 1)
 984                         fprintf(output, " (%d runs)", run_count);
 985                 fprintf(output, ":\n\n");
 986         }
 987
 988         if (no_aggr) {
 989                 list_for_each_entry(counter, &evsel_list->entries, node)
 990                         print_counter(counter);
 991         } else {
 992                 list_for_each_entry(counter, &evsel_list->entries, node)
 993                         print_counter_aggr(counter);
 994         }
 995
 996         if (!csv_output) {
 997                 if (!null_run)
 998                         fprintf(output, "\n");
 999                 fprintf(output, " %17.9f seconds time elapsed",
1000                                 avg_stats(&walltime_nsecs_stats)/1e9);
1001                 if (run_count > 1) {
1002                         fprintf(output, "                                        ");
1003                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1004                                         avg_stats(&walltime_nsecs_stats));
1005                 }
1006                 fprintf(output, "\n\n");
1007         }
1008 }
1009
1010 static volatile int signr = -1;
1011
1012 static void skip_signal(int signo)
1013 {
1014         if(child_pid == -1)
1015                 done = 1;
1016
1017         signr = signo;
1018 }
1019
1020 static void sig_atexit(void)
1021 {
1022         if (child_pid != -1)
1023                 kill(child_pid, SIGTERM);
1024
1025         if (signr == -1)
1026                 return;
1027
1028         signal(signr, SIG_DFL);
1029         kill(getpid(), signr);
1030 }
1031
1032 static const char * const stat_usage[] = {
1033         "perf stat [<options>] [<command>]",
1034         NULL
1035 };
1036
1037 static int stat__set_big_num(const struct option *opt __used,
1038                              const char *s __used, int unset)
1039 {
1040         big_num_opt = unset ? 0 : 1;
1041         return 0;
1042 }
1043
1044 static bool append_file;
1045
1046 static const struct option options[] = {
1047         OPT_CALLBACK('e', "event", &evsel_list, "event",
1048                      "event selector. use 'perf list' to list available events",
1049                      parse_events_option),
1050         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1051                      "event filter", parse_filter),
1052         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1053                     "child tasks do not inherit counters"),
1054         OPT_INTEGER('p', "pid", &target_pid,
1055                     "stat events on existing process id"),
1056         OPT_INTEGER('t', "tid", &target_tid,
1057                     "stat events on existing thread id"),
1058         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1059                     "system-wide collection from all CPUs"),
1060         OPT_BOOLEAN('g', "group", &group,
1061                     "put the counters into a counter group"),
1062         OPT_BOOLEAN('c', "scale", &scale,
1063                     "scale/normalize counters"),
1064         OPT_INCR('v', "verbose", &verbose,
1065                     "be more verbose (show counter open errors, etc)"),
1066         OPT_INTEGER('r', "repeat", &run_count,
1067                     "repeat command and print average + stddev (max: 100)"),
1068         OPT_BOOLEAN('n', "null", &null_run,
1069                     "null run - dont start any counters"),
1070         OPT_INCR('d', "detailed", &detailed_run,
1071                     "detailed run - start a lot of events"),
1072         OPT_BOOLEAN('S', "sync", &sync_run,
1073                     "call sync() before starting a run"),
1074         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1075                            "print large numbers with thousands\' separators",
1076                            stat__set_big_num),
1077         OPT_STRING('C', "cpu", &cpu_list, "cpu",
1078                     "list of cpus to monitor in system-wide"),
1079         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
1080                     "disable CPU count aggregation"),
1081         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1082                    "print counts with custom separator"),
1083         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1084                      "monitor event in cgroup name only",
1085                      parse_cgroups),
1086         OPT_STRING('o', "output", &output_name, "file",
1087                     "output file name"),
1088         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1089         OPT_INTEGER(0, "log-fd", &output_fd,
1090                     "log output to fd, instead of stderr"),
1091         OPT_END()
1092 };
1093
1094 /*
1095  * Add default attributes, if there were no attributes specified or
1096  * if -d/--detailed, -d -d or -d -d -d is used:
1097  */
1098 static int add_default_attributes(void)
1099 {
1100         struct perf_evsel *pos;
1101         size_t attr_nr = 0;
1102         size_t c;
1103
1104         /* Set attrs if no event is selected and !null_run: */
1105         if (null_run)
1106                 return 0;
1107
1108         if (!evsel_list->nr_entries) {
1109                 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) {
1110                         pos = perf_evsel__new(default_attrs + c, c + attr_nr);
1111                         if (pos == NULL)
1112                                 return -1;
1113                         perf_evlist__add(evsel_list, pos);
1114                 }
1115                 attr_nr += c;
1116         }
1117
1118         /* Detailed events get appended to the event list: */
1119
1120         if (detailed_run <  1)
1121                 return 0;
1122
1123         /* Append detailed run extra attributes: */
1124         for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) {
1125                 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr);
1126                 if (pos == NULL)
1127                         return -1;
1128                 perf_evlist__add(evsel_list, pos);
1129         }
1130         attr_nr += c;
1131
1132         if (detailed_run < 2)
1133                 return 0;
1134
1135         /* Append very detailed run extra attributes: */
1136         for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) {
1137                 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr);
1138                 if (pos == NULL)
1139                         return -1;
1140                 perf_evlist__add(evsel_list, pos);
1141         }
1142
1143         if (detailed_run < 3)
1144                 return 0;
1145
1146         /* Append very, very detailed run extra attributes: */
1147         for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) {
1148                 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr);
1149                 if (pos == NULL)
1150                         return -1;
1151                 perf_evlist__add(evsel_list, pos);
1152         }
1153
1154
1155         return 0;
1156 }
1157
1158 int cmd_stat(int argc, const char **argv, const char *prefix __used)
1159 {
1160         struct perf_evsel *pos;
1161         int status = -ENOMEM;
1162         const char *mode;
1163
1164         setlocale(LC_ALL, "");
1165
1166         evsel_list = perf_evlist__new(NULL, NULL);
1167         if (evsel_list == NULL)
1168                 return -ENOMEM;
1169
1170         argc = parse_options(argc, argv, options, stat_usage,
1171                 PARSE_OPT_STOP_AT_NON_OPTION);
1172
1173         output = stderr;
1174         if (output_name && strcmp(output_name, "-"))
1175                 output = NULL;
1176
1177         if (output_name && output_fd) {
1178                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1179                 usage_with_options(stat_usage, options);
1180         }
1181         if (!output) {
1182                 struct timespec tm;
1183                 mode = append_file ? "a" : "w";
1184
1185                 output = fopen(output_name, mode);
1186                 if (!output) {
1187                         perror("failed to create output file");
1188                         exit(-1);
1189                 }
1190                 clock_gettime(CLOCK_REALTIME, &tm);
1191                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1192         } else if (output_fd != 2) {
1193                 mode = append_file ? "a" : "w";
1194                 output = fdopen(output_fd, mode);
1195                 if (!output) {
1196                         perror("Failed opening logfd");
1197                         return -errno;
1198                 }
1199         }
1200
1201         if (csv_sep)
1202                 csv_output = true;
1203         else
1204                 csv_sep = DEFAULT_SEPARATOR;
1205
1206         /*
1207          * let the spreadsheet do the pretty-printing
1208          */
1209         if (csv_output) {
1210                 /* User explicitely passed -B? */
1211                 if (big_num_opt == 1) {
1212                         fprintf(stderr, "-B option not supported with -x\n");
1213                         usage_with_options(stat_usage, options);
1214                 } else /* Nope, so disable big number formatting */
1215                         big_num = false;
1216         } else if (big_num_opt == 0) /* User passed --no-big-num */
1217                 big_num = false;
1218
1219         if (!argc && target_pid == -1 && target_tid == -1)
1220                 usage_with_options(stat_usage, options);
1221         if (run_count <= 0)
1222                 usage_with_options(stat_usage, options);
1223
1224         /* no_aggr, cgroup are for system-wide only */
1225         if ((no_aggr || nr_cgroups) && !system_wide) {
1226                 fprintf(stderr, "both cgroup and no-aggregation "
1227                         "modes only available in system-wide mode\n");
1228
1229                 usage_with_options(stat_usage, options);
1230         }
1231
1232         if (add_default_attributes())
1233                 goto out;
1234
1235         if (target_pid != -1)
1236                 target_tid = target_pid;
1237
1238         evsel_list->threads = thread_map__new(target_pid, target_tid);
1239         if (evsel_list->threads == NULL) {
1240                 pr_err("Problems finding threads of monitor\n");
1241                 usage_with_options(stat_usage, options);
1242         }
1243
1244         if (system_wide)
1245                 evsel_list->cpus = cpu_map__new(cpu_list);
1246         else
1247                 evsel_list->cpus = cpu_map__dummy_new();
1248
1249         if (evsel_list->cpus == NULL) {
1250                 perror("failed to parse CPUs map");
1251                 usage_with_options(stat_usage, options);
1252                 return -1;
1253         }
1254
1255         list_for_each_entry(pos, &evsel_list->entries, node) {
1256                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1257                     perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
1258                     perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
1259                         goto out_free_fd;
1260         }
1261
1262         /*
1263          * We dont want to block the signals - that would cause
1264          * child tasks to inherit that and Ctrl-C would not work.
1265          * What we want is for Ctrl-C to work in the exec()-ed
1266          * task, but being ignored by perf stat itself:
1267          */
1268         atexit(sig_atexit);
1269         signal(SIGINT,  skip_signal);
1270         signal(SIGALRM, skip_signal);
1271         signal(SIGABRT, skip_signal);
1272
1273         status = 0;
1274         for (run_idx = 0; run_idx < run_count; run_idx++) {
1275                 if (run_count != 1 && verbose)
1276                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1277                                 run_idx + 1);
1278
1279                 if (sync_run)
1280                         sync();
1281
1282                 status = run_perf_stat(argc, argv);
1283         }
1284
1285         if (status != -1)
1286                 print_stat(argc, argv);
1287 out_free_fd:
1288         list_for_each_entry(pos, &evsel_list->entries, node)
1289                 perf_evsel__free_stat_priv(pos);
1290         perf_evlist__delete_maps(evsel_list);
1291 out:
1292         perf_evlist__delete(evsel_list);
1293         return status;
1294 }