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/cgroup.h"
  47 #include "util/util.h"
  48 #include "util/parse-options.h"
  49 #include "util/parse-events.h"
  50 #include "util/pmu.h"
  51 #include "util/event.h"
  52 #include "util/evlist.h"
  53 #include "util/evsel.h"
  54 #include "util/debug.h"
  55 #include "util/color.h"
  56 #include "util/stat.h"
  57 #include "util/header.h"
  58 #include "util/cpumap.h"
  59 #include "util/thread.h"
  60 #include "util/thread_map.h"
  61
  62 #include <stdlib.h>
  63 #include <sys/prctl.h>
  64 #include <locale.h>
  65
  66 #define DEFAULT_SEPARATOR       " "
  67 #define CNTR_NOT_SUPPORTED      "<not supported>"
  68 #define CNTR_NOT_COUNTED        "<not counted>"
  69
  70 static void print_counters(struct timespec *ts, int argc, const char **argv);
  71
  72 /* Default events used for perf stat -T */
  73 static const char *transaction_attrs = {
  74         "task-clock,"
  75         "{"
  76         "instructions,"
  77         "cycles,"
  78         "cpu/cycles-t/,"
  79         "cpu/tx-start/,"
  80         "cpu/el-start/,"
  81         "cpu/cycles-ct/"
  82         "}"
  83 };
  84
  85 /* More limited version when the CPU does not have all events. */
  86 static const char * transaction_limited_attrs = {
  87         "task-clock,"
  88         "{"
  89         "instructions,"
  90         "cycles,"
  91         "cpu/cycles-t/,"
  92         "cpu/tx-start/"
  93         "}"
  94 };
  95
  96 static struct perf_evlist       *evsel_list;
  97
  98 static struct target target = {
  99         .uid    = UINT_MAX,
 100 };
 101
 102 static int                      run_count                       =  1;
 103 static bool                     no_inherit                      = false;
 104 static bool                     scale                           =  true;
 105 static enum aggr_mode           aggr_mode                       = AGGR_GLOBAL;
 106 static volatile pid_t           child_pid                       = -1;
 107 static bool                     null_run                        =  false;
 108 static int                      detailed_run                    =  0;
 109 static bool                     transaction_run;
 110 static bool                     big_num                         =  true;
 111 static int                      big_num_opt                     =  -1;
 112 static const char               *csv_sep                        = NULL;
 113 static bool                     csv_output                      = false;
 114 static bool                     group                           = false;
 115 static FILE                     *output                         = NULL;
 116 static const char               *pre_cmd                        = NULL;
 117 static const char               *post_cmd                       = NULL;
 118 static bool                     sync_run                        = false;
 119 static unsigned int             interval                        = 0;
 120 static unsigned int             initial_delay                   = 0;
 121 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 122 static bool                     forever                         = false;
 123 static struct timespec          ref_time;
 124 static struct cpu_map           *aggr_map;
 125 static int                      (*aggr_get_id)(struct cpu_map *m, int cpu);
 126
 127 static volatile int done = 0;
 128
 129 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 130                                  struct timespec *b)
 131 {
 132         r->tv_sec = a->tv_sec - b->tv_sec;
 133         if (a->tv_nsec < b->tv_nsec) {
 134                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 135                 r->tv_sec--;
 136         } else {
 137                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 138         }
 139 }
 140
 141 static void perf_stat__reset_stats(void)
 142 {
 143         perf_evlist__reset_stats(evsel_list);
 144         perf_stat__reset_shadow_stats();
 145 }
 146
 147 static int create_perf_stat_counter(struct perf_evsel *evsel)
 148 {
 149         struct perf_event_attr *attr = &evsel->attr;
 150
 151         if (scale)
 152                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 153                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 154
 155         attr->inherit = !no_inherit;
 156
 157         if (target__has_cpu(&target))
 158                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 159
 160         if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
 161                 attr->disabled = 1;
 162                 if (!initial_delay)
 163                         attr->enable_on_exec = 1;
 164         }
 165
 166         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 167 }
 168
 169 /*
 170  * Does the counter have nsecs as a unit?
 171  */
 172 static inline int nsec_counter(struct perf_evsel *evsel)
 173 {
 174         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 175             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 176                 return 1;
 177
 178         return 0;
 179 }
 180
 181 static void zero_per_pkg(struct perf_evsel *counter)
 182 {
 183         if (counter->per_pkg_mask)
 184                 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
 185 }
 186
 187 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
 188 {
 189         unsigned long *mask = counter->per_pkg_mask;
 190         struct cpu_map *cpus = perf_evsel__cpus(counter);
 191         int s;
 192
 193         *skip = false;
 194
 195         if (!counter->per_pkg)
 196                 return 0;
 197
 198         if (cpu_map__empty(cpus))
 199                 return 0;
 200
 201         if (!mask) {
 202                 mask = zalloc(MAX_NR_CPUS);
 203                 if (!mask)
 204                         return -ENOMEM;
 205
 206                 counter->per_pkg_mask = mask;
 207         }
 208
 209         s = cpu_map__get_socket(cpus, cpu);
 210         if (s < 0)
 211                 return -1;
 212
 213         *skip = test_and_set_bit(s, mask) == 1;
 214         return 0;
 215 }
 216
 217 static int
 218 process_counter_values(struct perf_evsel *evsel, int cpu, int thread,
 219                        struct perf_counts_values *count)
 220 {
 221         struct perf_counts_values *aggr = &evsel->counts->aggr;
 222         static struct perf_counts_values zero;
 223         bool skip = false;
 224
 225         if (check_per_pkg(evsel, cpu, &skip)) {
 226                 pr_err("failed to read per-pkg counter\n");
 227                 return -1;
 228         }
 229
 230         if (skip)
 231                 count = &zero;
 232
 233         switch (aggr_mode) {
 234         case AGGR_THREAD:
 235         case AGGR_CORE:
 236         case AGGR_SOCKET:
 237         case AGGR_NONE:
 238                 if (!evsel->snapshot)
 239                         perf_evsel__compute_deltas(evsel, cpu, thread, count);
 240                 perf_counts_values__scale(count, scale, NULL);
 241                 if (aggr_mode == AGGR_NONE)
 242                         perf_stat__update_shadow_stats(evsel, count->values, cpu);
 243                 break;
 244         case AGGR_GLOBAL:
 245                 aggr->val += count->val;
 246                 if (scale) {
 247                         aggr->ena += count->ena;
 248                         aggr->run += count->run;
 249                 }
 250         default:
 251                 break;
 252         }
 253
 254         return 0;
 255 }
 256
 257 static int process_counter_maps(struct perf_evsel *counter)
 258 {
 259         int nthreads = thread_map__nr(counter->threads);
 260         int ncpus = perf_evsel__nr_cpus(counter);
 261         int cpu, thread;
 262
 263         if (counter->system_wide)
 264                 nthreads = 1;
 265
 266         for (thread = 0; thread < nthreads; thread++) {
 267                 for (cpu = 0; cpu < ncpus; cpu++) {
 268                         if (process_counter_values(counter, cpu, thread,
 269                                                    perf_counts(counter->counts, cpu, thread)))
 270                                 return -1;
 271                 }
 272         }
 273
 274         return 0;
 275 }
 276
 277 static int process_counter(struct perf_evsel *counter)
 278 {
 279         struct perf_counts_values *aggr = &counter->counts->aggr;
 280         struct perf_stat *ps = counter->priv;
 281         u64 *count = counter->counts->aggr.values;
 282         int i, ret;
 283
 284         aggr->val = aggr->ena = aggr->run = 0;
 285         init_stats(ps->res_stats);
 286
 287         if (counter->per_pkg)
 288                 zero_per_pkg(counter);
 289
 290         ret = process_counter_maps(counter);
 291         if (ret)
 292                 return ret;
 293
 294         if (aggr_mode != AGGR_GLOBAL)
 295                 return 0;
 296
 297         if (!counter->snapshot)
 298                 perf_evsel__compute_deltas(counter, -1, -1, aggr);
 299         perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
 300
 301         for (i = 0; i < 3; i++)
 302                 update_stats(&ps->res_stats[i], count[i]);
 303
 304         if (verbose) {
 305                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 306                         perf_evsel__name(counter), count[0], count[1], count[2]);
 307         }
 308
 309         /*
 310          * Save the full runtime - to allow normalization during printout:
 311          */
 312         perf_stat__update_shadow_stats(counter, count, 0);
 313
 314         return 0;
 315 }
 316
 317 /*
 318  * Read out the results of a single counter:
 319  * do not aggregate counts across CPUs in system-wide mode
 320  */
 321 static int read_counter(struct perf_evsel *counter)
 322 {
 323         int nthreads = thread_map__nr(evsel_list->threads);
 324         int ncpus = perf_evsel__nr_cpus(counter);
 325         int cpu, thread;
 326
 327         if (!counter->supported)
 328                 return -ENOENT;
 329
 330         if (counter->system_wide)
 331                 nthreads = 1;
 332
 333         for (thread = 0; thread < nthreads; thread++) {
 334                 for (cpu = 0; cpu < ncpus; cpu++) {
 335                         struct perf_counts_values *count;
 336
 337                         count = perf_counts(counter->counts, cpu, thread);
 338                         if (perf_evsel__read(counter, cpu, thread, count))
 339                                 return -1;
 340                 }
 341         }
 342
 343         return 0;
 344 }
 345
 346 static void read_counters(bool close_counters)
 347 {
 348         struct perf_evsel *counter;
 349
 350         evlist__for_each(evsel_list, counter) {
 351                 if (read_counter(counter))
 352                         pr_warning("failed to read counter %s\n", counter->name);
 353
 354                 if (process_counter(counter))
 355                         pr_warning("failed to process counter %s\n", counter->name);
 356
 357                 if (close_counters) {
 358                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 359                                              thread_map__nr(evsel_list->threads));
 360                 }
 361         }
 362 }
 363
 364 static void process_interval(void)
 365 {
 366         struct timespec ts, rs;
 367
 368         read_counters(false);
 369
 370         clock_gettime(CLOCK_MONOTONIC, &ts);
 371         diff_timespec(&rs, &ts, &ref_time);
 372
 373         print_counters(&rs, 0, NULL);
 374 }
 375
 376 static void handle_initial_delay(void)
 377 {
 378         struct perf_evsel *counter;
 379
 380         if (initial_delay) {
 381                 const int ncpus = cpu_map__nr(evsel_list->cpus),
 382                         nthreads = thread_map__nr(evsel_list->threads);
 383
 384                 usleep(initial_delay * 1000);
 385                 evlist__for_each(evsel_list, counter)
 386                         perf_evsel__enable(counter, ncpus, nthreads);
 387         }
 388 }
 389
 390 static volatile int workload_exec_errno;
 391
 392 /*
 393  * perf_evlist__prepare_workload will send a SIGUSR1
 394  * if the fork fails, since we asked by setting its
 395  * want_signal to true.
 396  */
 397 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 398                                         void *ucontext __maybe_unused)
 399 {
 400         workload_exec_errno = info->si_value.sival_int;
 401 }
 402
 403 static int __run_perf_stat(int argc, const char **argv)
 404 {
 405         char msg[512];
 406         unsigned long long t0, t1;
 407         struct perf_evsel *counter;
 408         struct timespec ts;
 409         size_t l;
 410         int status = 0;
 411         const bool forks = (argc > 0);
 412
 413         if (interval) {
 414                 ts.tv_sec  = interval / 1000;
 415                 ts.tv_nsec = (interval % 1000) * 1000000;
 416         } else {
 417                 ts.tv_sec  = 1;
 418                 ts.tv_nsec = 0;
 419         }
 420
 421         if (forks) {
 422                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
 423                                                   workload_exec_failed_signal) < 0) {
 424                         perror("failed to prepare workload");
 425                         return -1;
 426                 }
 427                 child_pid = evsel_list->workload.pid;
 428         }
 429
 430         if (group)
 431                 perf_evlist__set_leader(evsel_list);
 432
 433         evlist__for_each(evsel_list, counter) {
 434                 if (create_perf_stat_counter(counter) < 0) {
 435                         /*
 436                          * PPC returns ENXIO for HW counters until 2.6.37
 437                          * (behavior changed with commit b0a873e).
 438                          */
 439                         if (errno == EINVAL || errno == ENOSYS ||
 440                             errno == ENOENT || errno == EOPNOTSUPP ||
 441                             errno == ENXIO) {
 442                                 if (verbose)
 443                                         ui__warning("%s event is not supported by the kernel.\n",
 444                                                     perf_evsel__name(counter));
 445                                 counter->supported = false;
 446
 447                                 if ((counter->leader != counter) ||
 448                                     !(counter->leader->nr_members > 1))
 449                                         continue;
 450                         }
 451
 452                         perf_evsel__open_strerror(counter, &target,
 453                                                   errno, msg, sizeof(msg));
 454                         ui__error("%s\n", msg);
 455
 456                         if (child_pid != -1)
 457                                 kill(child_pid, SIGTERM);
 458
 459                         return -1;
 460                 }
 461                 counter->supported = true;
 462
 463                 l = strlen(counter->unit);
 464                 if (l > unit_width)
 465                         unit_width = l;
 466         }
 467
 468         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 469                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 470                         counter->filter, perf_evsel__name(counter), errno,
 471                         strerror_r(errno, msg, sizeof(msg)));
 472                 return -1;
 473         }
 474
 475         /*
 476          * Enable counters and exec the command:
 477          */
 478         t0 = rdclock();
 479         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 480
 481         if (forks) {
 482                 perf_evlist__start_workload(evsel_list);
 483                 handle_initial_delay();
 484
 485                 if (interval) {
 486                         while (!waitpid(child_pid, &status, WNOHANG)) {
 487                                 nanosleep(&ts, NULL);
 488                                 process_interval();
 489                         }
 490                 }
 491                 wait(&status);
 492
 493                 if (workload_exec_errno) {
 494                         const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 495                         pr_err("Workload failed: %s\n", emsg);
 496                         return -1;
 497                 }
 498
 499                 if (WIFSIGNALED(status))
 500                         psignal(WTERMSIG(status), argv[0]);
 501         } else {
 502                 handle_initial_delay();
 503                 while (!done) {
 504                         nanosleep(&ts, NULL);
 505                         if (interval)
 506                                 process_interval();
 507                 }
 508         }
 509
 510         t1 = rdclock();
 511
 512         update_stats(&walltime_nsecs_stats, t1 - t0);
 513
 514         read_counters(true);
 515
 516         return WEXITSTATUS(status);
 517 }
 518
 519 static int run_perf_stat(int argc, const char **argv)
 520 {
 521         int ret;
 522
 523         if (pre_cmd) {
 524                 ret = system(pre_cmd);
 525                 if (ret)
 526                         return ret;
 527         }
 528
 529         if (sync_run)
 530                 sync();
 531
 532         ret = __run_perf_stat(argc, argv);
 533         if (ret)
 534                 return ret;
 535
 536         if (post_cmd) {
 537                 ret = system(post_cmd);
 538                 if (ret)
 539                         return ret;
 540         }
 541
 542         return ret;
 543 }
 544
 545 static void print_running(u64 run, u64 ena)
 546 {
 547         if (csv_output) {
 548                 fprintf(output, "%s%" PRIu64 "%s%.2f",
 549                                         csv_sep,
 550                                         run,
 551                                         csv_sep,
 552                                         ena ? 100.0 * run / ena : 100.0);
 553         } else if (run != ena) {
 554                 fprintf(output, "  (%.2f%%)", 100.0 * run / ena);
 555         }
 556 }
 557
 558 static void print_noise_pct(double total, double avg)
 559 {
 560         double pct = rel_stddev_stats(total, avg);
 561
 562         if (csv_output)
 563                 fprintf(output, "%s%.2f%%", csv_sep, pct);
 564         else if (pct)
 565                 fprintf(output, "  ( +-%6.2f%% )", pct);
 566 }
 567
 568 static void print_noise(struct perf_evsel *evsel, double avg)
 569 {
 570         struct perf_stat *ps;
 571
 572         if (run_count == 1)
 573                 return;
 574
 575         ps = evsel->priv;
 576         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 577 }
 578
 579 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 580 {
 581         switch (aggr_mode) {
 582         case AGGR_CORE:
 583                 fprintf(output, "S%d-C%*d%s%*d%s",
 584                         cpu_map__id_to_socket(id),
 585                         csv_output ? 0 : -8,
 586                         cpu_map__id_to_cpu(id),
 587                         csv_sep,
 588                         csv_output ? 0 : 4,
 589                         nr,
 590                         csv_sep);
 591                 break;
 592         case AGGR_SOCKET:
 593                 fprintf(output, "S%*d%s%*d%s",
 594                         csv_output ? 0 : -5,
 595                         id,
 596                         csv_sep,
 597                         csv_output ? 0 : 4,
 598                         nr,
 599                         csv_sep);
 600                         break;
 601         case AGGR_NONE:
 602                 fprintf(output, "CPU%*d%s",
 603                         csv_output ? 0 : -4,
 604                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 605                 break;
 606         case AGGR_THREAD:
 607                 fprintf(output, "%*s-%*d%s",
 608                         csv_output ? 0 : 16,
 609                         thread_map__comm(evsel->threads, id),
 610                         csv_output ? 0 : -8,
 611                         thread_map__pid(evsel->threads, id),
 612                         csv_sep);
 613                 break;
 614         case AGGR_GLOBAL:
 615         default:
 616                 break;
 617         }
 618 }
 619
 620 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 621 {
 622         double msecs = avg / 1e6;
 623         const char *fmt_v, *fmt_n;
 624         char name[25];
 625
 626         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
 627         fmt_n = csv_output ? "%s" : "%-25s";
 628
 629         aggr_printout(evsel, id, nr);
 630
 631         scnprintf(name, sizeof(name), "%s%s",
 632                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
 633
 634         fprintf(output, fmt_v, msecs, csv_sep);
 635
 636         if (csv_output)
 637                 fprintf(output, "%s%s", evsel->unit, csv_sep);
 638         else
 639                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
 640
 641         fprintf(output, fmt_n, name);
 642
 643         if (evsel->cgrp)
 644                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 645
 646         if (csv_output || interval)
 647                 return;
 648
 649         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 650                 fprintf(output, " # %8.3f CPUs utilized          ",
 651                         avg / avg_stats(&walltime_nsecs_stats));
 652         else
 653                 fprintf(output, "                                   ");
 654 }
 655
 656 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 657 {
 658         double sc =  evsel->scale;
 659         const char *fmt;
 660         int cpu = cpu_map__id_to_cpu(id);
 661
 662         if (csv_output) {
 663                 fmt = sc != 1.0 ?  "%.2f%s" : "%.0f%s";
 664         } else {
 665                 if (big_num)
 666                         fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
 667                 else
 668                         fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
 669         }
 670
 671         aggr_printout(evsel, id, nr);
 672
 673         if (aggr_mode == AGGR_GLOBAL)
 674                 cpu = 0;
 675
 676         fprintf(output, fmt, avg, csv_sep);
 677
 678         if (evsel->unit)
 679                 fprintf(output, "%-*s%s",
 680                         csv_output ? 0 : unit_width,
 681                         evsel->unit, csv_sep);
 682
 683         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
 684
 685         if (evsel->cgrp)
 686                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 687
 688         if (csv_output || interval)
 689                 return;
 690
 691         perf_stat__print_shadow_stats(output, evsel, avg, cpu, aggr_mode);
 692 }
 693
 694 static void print_aggr(char *prefix)
 695 {
 696         struct perf_evsel *counter;
 697         int cpu, cpu2, s, s2, id, nr;
 698         double uval;
 699         u64 ena, run, val;
 700
 701         if (!(aggr_map || aggr_get_id))
 702                 return;
 703
 704         for (s = 0; s < aggr_map->nr; s++) {
 705                 id = aggr_map->map[s];
 706                 evlist__for_each(evsel_list, counter) {
 707                         val = ena = run = 0;
 708                         nr = 0;
 709                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 710                                 cpu2 = perf_evsel__cpus(counter)->map[cpu];
 711                                 s2 = aggr_get_id(evsel_list->cpus, cpu2);
 712                                 if (s2 != id)
 713                                         continue;
 714                                 val += perf_counts(counter->counts, cpu, 0)->val;
 715                                 ena += perf_counts(counter->counts, cpu, 0)->ena;
 716                                 run += perf_counts(counter->counts, cpu, 0)->run;
 717                                 nr++;
 718                         }
 719                         if (prefix)
 720                                 fprintf(output, "%s", prefix);
 721
 722                         if (run == 0 || ena == 0) {
 723                                 aggr_printout(counter, id, nr);
 724
 725                                 fprintf(output, "%*s%s",
 726                                         csv_output ? 0 : 18,
 727                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 728                                         csv_sep);
 729
 730                                 fprintf(output, "%-*s%s",
 731                                         csv_output ? 0 : unit_width,
 732                                         counter->unit, csv_sep);
 733
 734                                 fprintf(output, "%*s",
 735                                         csv_output ? 0 : -25,
 736                                         perf_evsel__name(counter));
 737
 738                                 if (counter->cgrp)
 739                                         fprintf(output, "%s%s",
 740                                                 csv_sep, counter->cgrp->name);
 741
 742                                 print_running(run, ena);
 743                                 fputc('\n', output);
 744                                 continue;
 745                         }
 746                         uval = val * counter->scale;
 747
 748                         if (nsec_counter(counter))
 749                                 nsec_printout(id, nr, counter, uval);
 750                         else
 751                                 abs_printout(id, nr, counter, uval);
 752
 753                         if (!csv_output)
 754                                 print_noise(counter, 1.0);
 755
 756                         print_running(run, ena);
 757                         fputc('\n', output);
 758                 }
 759         }
 760 }
 761
 762 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
 763 {
 764         int nthreads = thread_map__nr(counter->threads);
 765         int ncpus = cpu_map__nr(counter->cpus);
 766         int cpu, thread;
 767         double uval;
 768
 769         for (thread = 0; thread < nthreads; thread++) {
 770                 u64 ena = 0, run = 0, val = 0;
 771
 772                 for (cpu = 0; cpu < ncpus; cpu++) {
 773                         val += perf_counts(counter->counts, cpu, thread)->val;
 774                         ena += perf_counts(counter->counts, cpu, thread)->ena;
 775                         run += perf_counts(counter->counts, cpu, thread)->run;
 776                 }
 777
 778                 if (prefix)
 779                         fprintf(output, "%s", prefix);
 780
 781                 uval = val * counter->scale;
 782
 783                 if (nsec_counter(counter))
 784                         nsec_printout(thread, 0, counter, uval);
 785                 else
 786                         abs_printout(thread, 0, counter, uval);
 787
 788                 if (!csv_output)
 789                         print_noise(counter, 1.0);
 790
 791                 print_running(run, ena);
 792                 fputc('\n', output);
 793         }
 794 }
 795
 796 /*
 797  * Print out the results of a single counter:
 798  * aggregated counts in system-wide mode
 799  */
 800 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
 801 {
 802         struct perf_stat *ps = counter->priv;
 803         double avg = avg_stats(&ps->res_stats[0]);
 804         int scaled = counter->counts->scaled;
 805         double uval;
 806         double avg_enabled, avg_running;
 807
 808         avg_enabled = avg_stats(&ps->res_stats[1]);
 809         avg_running = avg_stats(&ps->res_stats[2]);
 810
 811         if (prefix)
 812                 fprintf(output, "%s", prefix);
 813
 814         if (scaled == -1 || !counter->supported) {
 815                 fprintf(output, "%*s%s",
 816                         csv_output ? 0 : 18,
 817                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 818                         csv_sep);
 819                 fprintf(output, "%-*s%s",
 820                         csv_output ? 0 : unit_width,
 821                         counter->unit, csv_sep);
 822                 fprintf(output, "%*s",
 823                         csv_output ? 0 : -25,
 824                         perf_evsel__name(counter));
 825
 826                 if (counter->cgrp)
 827                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
 828
 829                 print_running(avg_running, avg_enabled);
 830                 fputc('\n', output);
 831                 return;
 832         }
 833
 834         uval = avg * counter->scale;
 835
 836         if (nsec_counter(counter))
 837                 nsec_printout(-1, 0, counter, uval);
 838         else
 839                 abs_printout(-1, 0, counter, uval);
 840
 841         print_noise(counter, avg);
 842
 843         print_running(avg_running, avg_enabled);
 844         fprintf(output, "\n");
 845 }
 846
 847 /*
 848  * Print out the results of a single counter:
 849  * does not use aggregated count in system-wide
 850  */
 851 static void print_counter(struct perf_evsel *counter, char *prefix)
 852 {
 853         u64 ena, run, val;
 854         double uval;
 855         int cpu;
 856
 857         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 858                 val = perf_counts(counter->counts, cpu, 0)->val;
 859                 ena = perf_counts(counter->counts, cpu, 0)->ena;
 860                 run = perf_counts(counter->counts, cpu, 0)->run;
 861
 862                 if (prefix)
 863                         fprintf(output, "%s", prefix);
 864
 865                 if (run == 0 || ena == 0) {
 866                         fprintf(output, "CPU%*d%s%*s%s",
 867                                 csv_output ? 0 : -4,
 868                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
 869                                 csv_output ? 0 : 18,
 870                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 871                                 csv_sep);
 872
 873                                 fprintf(output, "%-*s%s",
 874                                         csv_output ? 0 : unit_width,
 875                                         counter->unit, csv_sep);
 876
 877                                 fprintf(output, "%*s",
 878                                         csv_output ? 0 : -25,
 879                                         perf_evsel__name(counter));
 880
 881                         if (counter->cgrp)
 882                                 fprintf(output, "%s%s",
 883                                         csv_sep, counter->cgrp->name);
 884
 885                         print_running(run, ena);
 886                         fputc('\n', output);
 887                         continue;
 888                 }
 889
 890                 uval = val * counter->scale;
 891
 892                 if (nsec_counter(counter))
 893                         nsec_printout(cpu, 0, counter, uval);
 894                 else
 895                         abs_printout(cpu, 0, counter, uval);
 896
 897                 if (!csv_output)
 898                         print_noise(counter, 1.0);
 899                 print_running(run, ena);
 900
 901                 fputc('\n', output);
 902         }
 903 }
 904
 905 static void print_interval(char *prefix, struct timespec *ts)
 906 {
 907         static int num_print_interval;
 908
 909         sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
 910
 911         if (num_print_interval == 0 && !csv_output) {
 912                 switch (aggr_mode) {
 913                 case AGGR_SOCKET:
 914                         fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
 915                         break;
 916                 case AGGR_CORE:
 917                         fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
 918                         break;
 919                 case AGGR_NONE:
 920                         fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
 921                         break;
 922                 case AGGR_THREAD:
 923                         fprintf(output, "#           time             comm-pid                  counts %*s events\n", unit_width, "unit");
 924                         break;
 925                 case AGGR_GLOBAL:
 926                 default:
 927                         fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
 928                 }
 929         }
 930
 931         if (++num_print_interval == 25)
 932                 num_print_interval = 0;
 933 }
 934
 935 static void print_header(int argc, const char **argv)
 936 {
 937         int i;
 938
 939         fflush(stdout);
 940
 941         if (!csv_output) {
 942                 fprintf(output, "\n");
 943                 fprintf(output, " Performance counter stats for ");
 944                 if (target.system_wide)
 945                         fprintf(output, "\'system wide");
 946                 else if (target.cpu_list)
 947                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
 948                 else if (!target__has_task(&target)) {
 949                         fprintf(output, "\'%s", argv[0]);
 950                         for (i = 1; i < argc; i++)
 951                                 fprintf(output, " %s", argv[i]);
 952                 } else if (target.pid)
 953                         fprintf(output, "process id \'%s", target.pid);
 954                 else
 955                         fprintf(output, "thread id \'%s", target.tid);
 956
 957                 fprintf(output, "\'");
 958                 if (run_count > 1)
 959                         fprintf(output, " (%d runs)", run_count);
 960                 fprintf(output, ":\n\n");
 961         }
 962 }
 963
 964 static void print_footer(void)
 965 {
 966         if (!null_run)
 967                 fprintf(output, "\n");
 968         fprintf(output, " %17.9f seconds time elapsed",
 969                         avg_stats(&walltime_nsecs_stats)/1e9);
 970         if (run_count > 1) {
 971                 fprintf(output, "                                        ");
 972                 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
 973                                 avg_stats(&walltime_nsecs_stats));
 974         }
 975         fprintf(output, "\n\n");
 976 }
 977
 978 static void print_counters(struct timespec *ts, int argc, const char **argv)
 979 {
 980         struct perf_evsel *counter;
 981         char buf[64], *prefix = NULL;
 982
 983         if (interval)
 984                 print_interval(prefix = buf, ts);
 985         else
 986                 print_header(argc, argv);
 987
 988         switch (aggr_mode) {
 989         case AGGR_CORE:
 990         case AGGR_SOCKET:
 991                 print_aggr(prefix);
 992                 break;
 993         case AGGR_THREAD:
 994                 evlist__for_each(evsel_list, counter)
 995                         print_aggr_thread(counter, prefix);
 996                 break;
 997         case AGGR_GLOBAL:
 998                 evlist__for_each(evsel_list, counter)
 999                         print_counter_aggr(counter, prefix);
1000                 break;
1001         case AGGR_NONE:
1002                 evlist__for_each(evsel_list, counter)
1003                         print_counter(counter, prefix);
1004                 break;
1005         default:
1006                 break;
1007         }
1008
1009         if (!interval && !csv_output)
1010                 print_footer();
1011
1012         fflush(output);
1013 }
1014
1015 static volatile int signr = -1;
1016
1017 static void skip_signal(int signo)
1018 {
1019         if ((child_pid == -1) || interval)
1020                 done = 1;
1021
1022         signr = signo;
1023         /*
1024          * render child_pid harmless
1025          * won't send SIGTERM to a random
1026          * process in case of race condition
1027          * and fast PID recycling
1028          */
1029         child_pid = -1;
1030 }
1031
1032 static void sig_atexit(void)
1033 {
1034         sigset_t set, oset;
1035
1036         /*
1037          * avoid race condition with SIGCHLD handler
1038          * in skip_signal() which is modifying child_pid
1039          * goal is to avoid send SIGTERM to a random
1040          * process
1041          */
1042         sigemptyset(&set);
1043         sigaddset(&set, SIGCHLD);
1044         sigprocmask(SIG_BLOCK, &set, &oset);
1045
1046         if (child_pid != -1)
1047                 kill(child_pid, SIGTERM);
1048
1049         sigprocmask(SIG_SETMASK, &oset, NULL);
1050
1051         if (signr == -1)
1052                 return;
1053
1054         signal(signr, SIG_DFL);
1055         kill(getpid(), signr);
1056 }
1057
1058 static int stat__set_big_num(const struct option *opt __maybe_unused,
1059                              const char *s __maybe_unused, int unset)
1060 {
1061         big_num_opt = unset ? 0 : 1;
1062         return 0;
1063 }
1064
1065 static int perf_stat_init_aggr_mode(void)
1066 {
1067         switch (aggr_mode) {
1068         case AGGR_SOCKET:
1069                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1070                         perror("cannot build socket map");
1071                         return -1;
1072                 }
1073                 aggr_get_id = cpu_map__get_socket;
1074                 break;
1075         case AGGR_CORE:
1076                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1077                         perror("cannot build core map");
1078                         return -1;
1079                 }
1080                 aggr_get_id = cpu_map__get_core;
1081                 break;
1082         case AGGR_NONE:
1083         case AGGR_GLOBAL:
1084         case AGGR_THREAD:
1085         default:
1086                 break;
1087         }
1088         return 0;
1089 }
1090
1091 /*
1092  * Add default attributes, if there were no attributes specified or
1093  * if -d/--detailed, -d -d or -d -d -d is used:
1094  */
1095 static int add_default_attributes(void)
1096 {
1097         struct perf_event_attr default_attrs[] = {
1098
1099   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1100   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1101   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1102   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1103
1104   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1105   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1106   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1107   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1108   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1109   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1110
1111 };
1112
1113 /*
1114  * Detailed stats (-d), covering the L1 and last level data caches:
1115  */
1116         struct perf_event_attr detailed_attrs[] = {
1117
1118   { .type = PERF_TYPE_HW_CACHE,
1119     .config =
1120          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1121         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1122         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1123
1124   { .type = PERF_TYPE_HW_CACHE,
1125     .config =
1126          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1127         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1128         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1129
1130   { .type = PERF_TYPE_HW_CACHE,
1131     .config =
1132          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1133         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1134         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1135
1136   { .type = PERF_TYPE_HW_CACHE,
1137     .config =
1138          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1139         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1140         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1141 };
1142
1143 /*
1144  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1145  */
1146         struct perf_event_attr very_detailed_attrs[] = {
1147
1148   { .type = PERF_TYPE_HW_CACHE,
1149     .config =
1150          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1151         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1152         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1153
1154   { .type = PERF_TYPE_HW_CACHE,
1155     .config =
1156          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1157         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1158         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1159
1160   { .type = PERF_TYPE_HW_CACHE,
1161     .config =
1162          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1163         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1164         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1165
1166   { .type = PERF_TYPE_HW_CACHE,
1167     .config =
1168          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1169         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1170         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1171
1172   { .type = PERF_TYPE_HW_CACHE,
1173     .config =
1174          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1175         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1176         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1177
1178   { .type = PERF_TYPE_HW_CACHE,
1179     .config =
1180          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1181         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1182         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1183
1184 };
1185
1186 /*
1187  * Very, very detailed stats (-d -d -d), adding prefetch events:
1188  */
1189         struct perf_event_attr very_very_detailed_attrs[] = {
1190
1191   { .type = PERF_TYPE_HW_CACHE,
1192     .config =
1193          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1194         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1195         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1196
1197   { .type = PERF_TYPE_HW_CACHE,
1198     .config =
1199          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1200         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1201         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1202 };
1203
1204         /* Set attrs if no event is selected and !null_run: */
1205         if (null_run)
1206                 return 0;
1207
1208         if (transaction_run) {
1209                 int err;
1210                 if (pmu_have_event("cpu", "cycles-ct") &&
1211                     pmu_have_event("cpu", "el-start"))
1212                         err = parse_events(evsel_list, transaction_attrs, NULL);
1213                 else
1214                         err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1215                 if (err) {
1216                         fprintf(stderr, "Cannot set up transaction events\n");
1217                         return -1;
1218                 }
1219                 return 0;
1220         }
1221
1222         if (!evsel_list->nr_entries) {
1223                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1224                         return -1;
1225         }
1226
1227         /* Detailed events get appended to the event list: */
1228
1229         if (detailed_run <  1)
1230                 return 0;
1231
1232         /* Append detailed run extra attributes: */
1233         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1234                 return -1;
1235
1236         if (detailed_run < 2)
1237                 return 0;
1238
1239         /* Append very detailed run extra attributes: */
1240         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1241                 return -1;
1242
1243         if (detailed_run < 3)
1244                 return 0;
1245
1246         /* Append very, very detailed run extra attributes: */
1247         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1248 }
1249
1250 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1251 {
1252         bool append_file = false;
1253         int output_fd = 0;
1254         const char *output_name = NULL;
1255         const struct option options[] = {
1256         OPT_BOOLEAN('T', "transaction", &transaction_run,
1257                     "hardware transaction statistics"),
1258         OPT_CALLBACK('e', "event", &evsel_list, "event",
1259                      "event selector. use 'perf list' to list available events",
1260                      parse_events_option),
1261         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1262                      "event filter", parse_filter),
1263         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1264                     "child tasks do not inherit counters"),
1265         OPT_STRING('p', "pid", &target.pid, "pid",
1266                    "stat events on existing process id"),
1267         OPT_STRING('t', "tid", &target.tid, "tid",
1268                    "stat events on existing thread id"),
1269         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1270                     "system-wide collection from all CPUs"),
1271         OPT_BOOLEAN('g', "group", &group,
1272                     "put the counters into a counter group"),
1273         OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1274         OPT_INCR('v', "verbose", &verbose,
1275                     "be more verbose (show counter open errors, etc)"),
1276         OPT_INTEGER('r', "repeat", &run_count,
1277                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1278         OPT_BOOLEAN('n', "null", &null_run,
1279                     "null run - dont start any counters"),
1280         OPT_INCR('d', "detailed", &detailed_run,
1281                     "detailed run - start a lot of events"),
1282         OPT_BOOLEAN('S', "sync", &sync_run,
1283                     "call sync() before starting a run"),
1284         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1285                            "print large numbers with thousands\' separators",
1286                            stat__set_big_num),
1287         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1288                     "list of cpus to monitor in system-wide"),
1289         OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1290                     "disable CPU count aggregation", AGGR_NONE),
1291         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1292                    "print counts with custom separator"),
1293         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1294                      "monitor event in cgroup name only", parse_cgroups),
1295         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1296         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1297         OPT_INTEGER(0, "log-fd", &output_fd,
1298                     "log output to fd, instead of stderr"),
1299         OPT_STRING(0, "pre", &pre_cmd, "command",
1300                         "command to run prior to the measured command"),
1301         OPT_STRING(0, "post", &post_cmd, "command",
1302                         "command to run after to the measured command"),
1303         OPT_UINTEGER('I', "interval-print", &interval,
1304                     "print counts at regular interval in ms (>= 100)"),
1305         OPT_SET_UINT(0, "per-socket", &aggr_mode,
1306                      "aggregate counts per processor socket", AGGR_SOCKET),
1307         OPT_SET_UINT(0, "per-core", &aggr_mode,
1308                      "aggregate counts per physical processor core", AGGR_CORE),
1309         OPT_SET_UINT(0, "per-thread", &aggr_mode,
1310                      "aggregate counts per thread", AGGR_THREAD),
1311         OPT_UINTEGER('D', "delay", &initial_delay,
1312                      "ms to wait before starting measurement after program start"),
1313         OPT_END()
1314         };
1315         const char * const stat_usage[] = {
1316                 "perf stat [<options>] [<command>]",
1317                 NULL
1318         };
1319         int status = -EINVAL, run_idx;
1320         const char *mode;
1321
1322         setlocale(LC_ALL, "");
1323
1324         evsel_list = perf_evlist__new();
1325         if (evsel_list == NULL)
1326                 return -ENOMEM;
1327
1328         argc = parse_options(argc, argv, options, stat_usage,
1329                 PARSE_OPT_STOP_AT_NON_OPTION);
1330
1331         output = stderr;
1332         if (output_name && strcmp(output_name, "-"))
1333                 output = NULL;
1334
1335         if (output_name && output_fd) {
1336                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1337                 parse_options_usage(stat_usage, options, "o", 1);
1338                 parse_options_usage(NULL, options, "log-fd", 0);
1339                 goto out;
1340         }
1341
1342         if (output_fd < 0) {
1343                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1344                 parse_options_usage(stat_usage, options, "log-fd", 0);
1345                 goto out;
1346         }
1347
1348         if (!output) {
1349                 struct timespec tm;
1350                 mode = append_file ? "a" : "w";
1351
1352                 output = fopen(output_name, mode);
1353                 if (!output) {
1354                         perror("failed to create output file");
1355                         return -1;
1356                 }
1357                 clock_gettime(CLOCK_REALTIME, &tm);
1358                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1359         } else if (output_fd > 0) {
1360                 mode = append_file ? "a" : "w";
1361                 output = fdopen(output_fd, mode);
1362                 if (!output) {
1363                         perror("Failed opening logfd");
1364                         return -errno;
1365                 }
1366         }
1367
1368         if (csv_sep) {
1369                 csv_output = true;
1370                 if (!strcmp(csv_sep, "\\t"))
1371                         csv_sep = "\t";
1372         } else
1373                 csv_sep = DEFAULT_SEPARATOR;
1374
1375         /*
1376          * let the spreadsheet do the pretty-printing
1377          */
1378         if (csv_output) {
1379                 /* User explicitly passed -B? */
1380                 if (big_num_opt == 1) {
1381                         fprintf(stderr, "-B option not supported with -x\n");
1382                         parse_options_usage(stat_usage, options, "B", 1);
1383                         parse_options_usage(NULL, options, "x", 1);
1384                         goto out;
1385                 } else /* Nope, so disable big number formatting */
1386                         big_num = false;
1387         } else if (big_num_opt == 0) /* User passed --no-big-num */
1388                 big_num = false;
1389
1390         if (!argc && target__none(&target))
1391                 usage_with_options(stat_usage, options);
1392
1393         if (run_count < 0) {
1394                 pr_err("Run count must be a positive number\n");
1395                 parse_options_usage(stat_usage, options, "r", 1);
1396                 goto out;
1397         } else if (run_count == 0) {
1398                 forever = true;
1399                 run_count = 1;
1400         }
1401
1402         if ((aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1403                 fprintf(stderr, "The --per-thread option is only available "
1404                         "when monitoring via -p -t options.\n");
1405                 parse_options_usage(NULL, options, "p", 1);
1406                 parse_options_usage(NULL, options, "t", 1);
1407                 goto out;
1408         }
1409
1410         /*
1411          * no_aggr, cgroup are for system-wide only
1412          * --per-thread is aggregated per thread, we dont mix it with cpu mode
1413          */
1414         if (((aggr_mode != AGGR_GLOBAL && aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1415             !target__has_cpu(&target)) {
1416                 fprintf(stderr, "both cgroup and no-aggregation "
1417                         "modes only available in system-wide mode\n");
1418
1419                 parse_options_usage(stat_usage, options, "G", 1);
1420                 parse_options_usage(NULL, options, "A", 1);
1421                 parse_options_usage(NULL, options, "a", 1);
1422                 goto out;
1423         }
1424
1425         if (add_default_attributes())
1426                 goto out;
1427
1428         target__validate(&target);
1429
1430         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1431                 if (target__has_task(&target)) {
1432                         pr_err("Problems finding threads of monitor\n");
1433                         parse_options_usage(stat_usage, options, "p", 1);
1434                         parse_options_usage(NULL, options, "t", 1);
1435                 } else if (target__has_cpu(&target)) {
1436                         perror("failed to parse CPUs map");
1437                         parse_options_usage(stat_usage, options, "C", 1);
1438                         parse_options_usage(NULL, options, "a", 1);
1439                 }
1440                 goto out;
1441         }
1442
1443         /*
1444          * Initialize thread_map with comm names,
1445          * so we could print it out on output.
1446          */
1447         if (aggr_mode == AGGR_THREAD)
1448                 thread_map__read_comms(evsel_list->threads);
1449
1450         if (interval && interval < 100) {
1451                 pr_err("print interval must be >= 100ms\n");
1452                 parse_options_usage(stat_usage, options, "I", 1);
1453                 goto out;
1454         }
1455
1456         if (perf_evlist__alloc_stats(evsel_list, interval))
1457                 goto out;
1458
1459         if (perf_stat_init_aggr_mode())
1460                 goto out;
1461
1462         /*
1463          * We dont want to block the signals - that would cause
1464          * child tasks to inherit that and Ctrl-C would not work.
1465          * What we want is for Ctrl-C to work in the exec()-ed
1466          * task, but being ignored by perf stat itself:
1467          */
1468         atexit(sig_atexit);
1469         if (!forever)
1470                 signal(SIGINT,  skip_signal);
1471         signal(SIGCHLD, skip_signal);
1472         signal(SIGALRM, skip_signal);
1473         signal(SIGABRT, skip_signal);
1474
1475         status = 0;
1476         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1477                 if (run_count != 1 && verbose)
1478                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1479                                 run_idx + 1);
1480
1481                 status = run_perf_stat(argc, argv);
1482                 if (forever && status != -1) {
1483                         print_counters(NULL, argc, argv);
1484                         perf_stat__reset_stats();
1485                 }
1486         }
1487
1488         if (!forever && status != -1 && !interval)
1489                 print_counters(NULL, argc, argv);
1490
1491         perf_evlist__free_stats(evsel_list);
1492 out:
1493         perf_evlist__delete(evsel_list);
1494         return status;
1495 }