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