tools/perf/builtin-stat.c

   1 /*
   2  * builtin-stat.c
   3  *
   4  * Builtin stat command: Give a precise performance counters summary
   5  * overview about any workload, CPU or specific PID.
   6  *
   7  * Sample output:
   8
   9    $ perf stat ./hackbench 10
  10
  11   Time: 0.118
  12
  13   Performance counter stats for './hackbench 10':
  14
  15        1708.761321 task-clock                #   11.037 CPUs utilized
  16             41,190 context-switches          #    0.024 M/sec
  17              6,735 CPU-migrations            #    0.004 M/sec
  18             17,318 page-faults               #    0.010 M/sec
  19      5,205,202,243 cycles                    #    3.046 GHz
  20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  22      2,603,501,247 instructions              #    0.50  insns per cycle
  23                                              #    1.48  stalled cycles per insn
  24        484,357,498 branches                  #  283.455 M/sec
  25          6,388,934 branch-misses             #    1.32% of all branches
  26
  27         0.154822978  seconds time elapsed
  28
  29  *
  30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  31  *
  32  * Improvements and fixes by:
  33  *
  34  *   Arjan van de Ven <arjan@linux.intel.com>
  35  *   Yanmin Zhang <yanmin.zhang@intel.com>
  36  *   Wu Fengguang <fengguang.wu@intel.com>
  37  *   Mike Galbraith <efault@gmx.de>
  38  *   Paul Mackerras <paulus@samba.org>
  39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  40  *
  41  * Released under the GPL v2. (and only v2, not any later version)
  42  */
  43
  44 #include "perf.h"
  45 #include "builtin.h"
  46 #include "util/util.h"
  47 #include "util/parse-options.h"
  48 #include "util/parse-events.h"
  49 #include "util/event.h"
  50 #include "util/evlist.h"
  51 #include "util/evsel.h"
  52 #include "util/debug.h"
  53 #include "util/color.h"
  54 #include "util/stat.h"
  55 #include "util/header.h"
  56 #include "util/cpumap.h"
  57 #include "util/thread.h"
  58 #include "util/thread_map.h"
  59
  60 #include <stdlib.h>
  61 #include <sys/prctl.h>
  62 #include <locale.h>
  63
  64 #define DEFAULT_SEPARATOR       " "
  65 #define CNTR_NOT_SUPPORTED      "<not supported>"
  66 #define CNTR_NOT_COUNTED        "<not counted>"
  67
  68 static void print_stat(int argc, const char **argv);
  69 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
  70 static void print_counter(struct perf_evsel *counter, char *prefix);
  71 static void print_aggr(char *prefix);
  72
  73 static struct perf_evlist       *evsel_list;
  74
  75 static struct perf_target       target = {
  76         .uid    = UINT_MAX,
  77 };
  78
  79 enum aggr_mode {
  80         AGGR_NONE,
  81         AGGR_GLOBAL,
  82         AGGR_SOCKET,
  83         AGGR_CORE,
  84 };
  85
  86 static int                      run_count                       =  1;
  87 static bool                     no_inherit                      = false;
  88 static bool                     scale                           =  true;
  89 static enum aggr_mode           aggr_mode                       = AGGR_GLOBAL;
  90 static volatile pid_t           child_pid                       = -1;
  91 static bool                     null_run                        =  false;
  92 static int                      detailed_run                    =  0;
  93 static bool                     big_num                         =  true;
  94 static int                      big_num_opt                     =  -1;
  95 static const char               *csv_sep                        = NULL;
  96 static bool                     csv_output                      = false;
  97 static bool                     group                           = false;
  98 static FILE                     *output                         = NULL;
  99 static const char               *pre_cmd                        = NULL;
 100 static const char               *post_cmd                       = NULL;
 101 static bool                     sync_run                        = false;
 102 static unsigned int             interval                        = 0;
 103 static unsigned int             initial_delay                   = 0;
 104 static bool                     forever                         = false;
 105 static struct timespec          ref_time;
 106 static struct cpu_map           *aggr_map;
 107 static int                      (*aggr_get_id)(struct cpu_map *m, int cpu);
 108
 109 static volatile int done = 0;
 110
 111 struct perf_stat {
 112         struct stats      res_stats[3];
 113 };
 114
 115 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 116                                  struct timespec *b)
 117 {
 118         r->tv_sec = a->tv_sec - b->tv_sec;
 119         if (a->tv_nsec < b->tv_nsec) {
 120                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 121                 r->tv_sec--;
 122         } else {
 123                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 124         }
 125 }
 126
 127 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
 128 {
 129         return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
 130 }
 131
 132 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
 133 {
 134         return perf_evsel__cpus(evsel)->nr;
 135 }
 136
 137 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 138 {
 139         memset(evsel->priv, 0, sizeof(struct perf_stat));
 140 }
 141
 142 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 143 {
 144         evsel->priv = zalloc(sizeof(struct perf_stat));
 145         return evsel->priv == NULL ? -ENOMEM : 0;
 146 }
 147
 148 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 149 {
 150         free(evsel->priv);
 151         evsel->priv = NULL;
 152 }
 153
 154 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
 155 {
 156         void *addr;
 157         size_t sz;
 158
 159         sz = sizeof(*evsel->counts) +
 160              (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
 161
 162         addr = zalloc(sz);
 163         if (!addr)
 164                 return -ENOMEM;
 165
 166         evsel->prev_raw_counts =  addr;
 167
 168         return 0;
 169 }
 170
 171 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
 172 {
 173         free(evsel->prev_raw_counts);
 174         evsel->prev_raw_counts = NULL;
 175 }
 176
 177 static void perf_evlist__free_stats(struct perf_evlist *evlist)
 178 {
 179         struct perf_evsel *evsel;
 180
 181         list_for_each_entry(evsel, &evlist->entries, node) {
 182                 perf_evsel__free_stat_priv(evsel);
 183                 perf_evsel__free_counts(evsel);
 184                 perf_evsel__free_prev_raw_counts(evsel);
 185         }
 186 }
 187
 188 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
 189 {
 190         struct perf_evsel *evsel;
 191
 192         list_for_each_entry(evsel, &evlist->entries, node) {
 193                 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 194                     perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
 195                     (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
 196                         goto out_free;
 197         }
 198
 199         return 0;
 200
 201 out_free:
 202         perf_evlist__free_stats(evlist);
 203         return -1;
 204 }
 205
 206 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
 207 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
 208 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
 209 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
 210 static struct stats runtime_branches_stats[MAX_NR_CPUS];
 211 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
 212 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
 213 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
 214 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
 215 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
 216 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
 217 static struct stats walltime_nsecs_stats;
 218
 219 static void perf_stat__reset_stats(struct perf_evlist *evlist)
 220 {
 221         struct perf_evsel *evsel;
 222
 223         list_for_each_entry(evsel, &evlist->entries, node) {
 224                 perf_evsel__reset_stat_priv(evsel);
 225                 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
 226         }
 227
 228         memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
 229         memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
 230         memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
 231         memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
 232         memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
 233         memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
 234         memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
 235         memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
 236         memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
 237         memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
 238         memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
 239         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_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 (perf_target__has_cpu(&target))
 253                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 254
 255         if (!perf_target__has_task(&target) &&
 256             perf_evsel__is_group_leader(evsel)) {
 257                 attr->disabled = 1;
 258                 if (!initial_delay)
 259                         attr->enable_on_exec = 1;
 260         }
 261
 262         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 263 }
 264
 265 /*
 266  * Does the counter have nsecs as a unit?
 267  */
 268 static inline int nsec_counter(struct perf_evsel *evsel)
 269 {
 270         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 271             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 272                 return 1;
 273
 274         return 0;
 275 }
 276
 277 /*
 278  * Update various tracking values we maintain to print
 279  * more semantic information such as miss/hit ratios,
 280  * instruction rates, etc:
 281  */
 282 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
 283 {
 284         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 285                 update_stats(&runtime_nsecs_stats[0], count[0]);
 286         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 287                 update_stats(&runtime_cycles_stats[0], count[0]);
 288         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
 289                 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
 290         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 291                 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
 292         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 293                 update_stats(&runtime_branches_stats[0], count[0]);
 294         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 295                 update_stats(&runtime_cacherefs_stats[0], count[0]);
 296         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
 297                 update_stats(&runtime_l1_dcache_stats[0], count[0]);
 298         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
 299                 update_stats(&runtime_l1_icache_stats[0], count[0]);
 300         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
 301                 update_stats(&runtime_ll_cache_stats[0], count[0]);
 302         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
 303                 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
 304         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
 305                 update_stats(&runtime_itlb_cache_stats[0], count[0]);
 306 }
 307
 308 /*
 309  * Read out the results of a single counter:
 310  * aggregate counts across CPUs in system-wide mode
 311  */
 312 static int read_counter_aggr(struct perf_evsel *counter)
 313 {
 314         struct perf_stat *ps = counter->priv;
 315         u64 *count = counter->counts->aggr.values;
 316         int i;
 317
 318         if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
 319                                thread_map__nr(evsel_list->threads), scale) < 0)
 320                 return -1;
 321
 322         for (i = 0; i < 3; i++)
 323                 update_stats(&ps->res_stats[i], count[i]);
 324
 325         if (verbose) {
 326                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 327                         perf_evsel__name(counter), count[0], count[1], count[2]);
 328         }
 329
 330         /*
 331          * Save the full runtime - to allow normalization during printout:
 332          */
 333         update_shadow_stats(counter, count);
 334
 335         return 0;
 336 }
 337
 338 /*
 339  * Read out the results of a single counter:
 340  * do not aggregate counts across CPUs in system-wide mode
 341  */
 342 static int read_counter(struct perf_evsel *counter)
 343 {
 344         u64 *count;
 345         int cpu;
 346
 347         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 348                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 349                         return -1;
 350
 351                 count = counter->counts->cpu[cpu].values;
 352
 353                 update_shadow_stats(counter, count);
 354         }
 355
 356         return 0;
 357 }
 358
 359 static void print_interval(void)
 360 {
 361         static int num_print_interval;
 362         struct perf_evsel *counter;
 363         struct perf_stat *ps;
 364         struct timespec ts, rs;
 365         char prefix[64];
 366
 367         if (aggr_mode == AGGR_GLOBAL) {
 368                 list_for_each_entry(counter, &evsel_list->entries, node) {
 369                         ps = counter->priv;
 370                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
 371                         read_counter_aggr(counter);
 372                 }
 373         } else  {
 374                 list_for_each_entry(counter, &evsel_list->entries, node) {
 375                         ps = counter->priv;
 376                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
 377                         read_counter(counter);
 378                 }
 379         }
 380
 381         clock_gettime(CLOCK_MONOTONIC, &ts);
 382         diff_timespec(&rs, &ts, &ref_time);
 383         sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
 384
 385         if (num_print_interval == 0 && !csv_output) {
 386                 switch (aggr_mode) {
 387                 case AGGR_SOCKET:
 388                         fprintf(output, "#           time socket cpus             counts events\n");
 389                         break;
 390                 case AGGR_CORE:
 391                         fprintf(output, "#           time core         cpus             counts events\n");
 392                         break;
 393                 case AGGR_NONE:
 394                         fprintf(output, "#           time CPU                 counts events\n");
 395                         break;
 396                 case AGGR_GLOBAL:
 397                 default:
 398                         fprintf(output, "#           time             counts events\n");
 399                 }
 400         }
 401
 402         if (++num_print_interval == 25)
 403                 num_print_interval = 0;
 404
 405         switch (aggr_mode) {
 406         case AGGR_CORE:
 407         case AGGR_SOCKET:
 408                 print_aggr(prefix);
 409                 break;
 410         case AGGR_NONE:
 411                 list_for_each_entry(counter, &evsel_list->entries, node)
 412                         print_counter(counter, prefix);
 413                 break;
 414         case AGGR_GLOBAL:
 415         default:
 416                 list_for_each_entry(counter, &evsel_list->entries, node)
 417                         print_counter_aggr(counter, prefix);
 418         }
 419 }
 420
 421 static void handle_initial_delay(void)
 422 {
 423         struct perf_evsel *counter;
 424
 425         if (initial_delay) {
 426                 const int ncpus = cpu_map__nr(evsel_list->cpus),
 427                         nthreads = thread_map__nr(evsel_list->threads);
 428
 429                 usleep(initial_delay * 1000);
 430                 list_for_each_entry(counter, &evsel_list->entries, node)
 431                         perf_evsel__enable(counter, ncpus, nthreads);
 432         }
 433 }
 434
 435 static int __run_perf_stat(int argc, const char **argv)
 436 {
 437         char msg[512];
 438         unsigned long long t0, t1;
 439         struct perf_evsel *counter;
 440         struct timespec ts;
 441         int status = 0;
 442         const bool forks = (argc > 0);
 443
 444         if (interval) {
 445                 ts.tv_sec  = interval / 1000;
 446                 ts.tv_nsec = (interval % 1000) * 1000000;
 447         } else {
 448                 ts.tv_sec  = 1;
 449                 ts.tv_nsec = 0;
 450         }
 451
 452         if (forks) {
 453                 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
 454                                                   false, false) < 0) {
 455                         perror("failed to prepare workload");
 456                         return -1;
 457                 }
 458         }
 459
 460         if (group)
 461                 perf_evlist__set_leader(evsel_list);
 462
 463         list_for_each_entry(counter, &evsel_list->entries, node) {
 464                 if (create_perf_stat_counter(counter) < 0) {
 465                         /*
 466                          * PPC returns ENXIO for HW counters until 2.6.37
 467                          * (behavior changed with commit b0a873e).
 468                          */
 469                         if (errno == EINVAL || errno == ENOSYS ||
 470                             errno == ENOENT || errno == EOPNOTSUPP ||
 471                             errno == ENXIO) {
 472                                 if (verbose)
 473                                         ui__warning("%s event is not supported by the kernel.\n",
 474                                                     perf_evsel__name(counter));
 475                                 counter->supported = false;
 476                                 continue;
 477                         }
 478
 479                         perf_evsel__open_strerror(counter, &target,
 480                                                   errno, msg, sizeof(msg));
 481                         ui__error("%s\n", msg);
 482
 483                         if (child_pid != -1)
 484                                 kill(child_pid, SIGTERM);
 485
 486                         return -1;
 487                 }
 488                 counter->supported = true;
 489         }
 490
 491         if (perf_evlist__apply_filters(evsel_list)) {
 492                 error("failed to set filter with %d (%s)\n", errno,
 493                         strerror(errno));
 494                 return -1;
 495         }
 496
 497         /*
 498          * Enable counters and exec the command:
 499          */
 500         t0 = rdclock();
 501         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 502
 503         if (forks) {
 504                 perf_evlist__start_workload(evsel_list);
 505                 handle_initial_delay();
 506
 507                 if (interval) {
 508                         while (!waitpid(child_pid, &status, WNOHANG)) {
 509                                 nanosleep(&ts, NULL);
 510                                 print_interval();
 511                         }
 512                 }
 513                 wait(&status);
 514                 if (WIFSIGNALED(status))
 515                         psignal(WTERMSIG(status), argv[0]);
 516         } else {
 517                 handle_initial_delay();
 518                 while (!done) {
 519                         nanosleep(&ts, NULL);
 520                         if (interval)
 521                                 print_interval();
 522                 }
 523         }
 524
 525         t1 = rdclock();
 526
 527         update_stats(&walltime_nsecs_stats, t1 - t0);
 528
 529         if (aggr_mode == AGGR_GLOBAL) {
 530                 list_for_each_entry(counter, &evsel_list->entries, node) {
 531                         read_counter_aggr(counter);
 532                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 533                                              thread_map__nr(evsel_list->threads));
 534                 }
 535         } else {
 536                 list_for_each_entry(counter, &evsel_list->entries, node) {
 537                         read_counter(counter);
 538                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
 539                 }
 540         }
 541
 542         return WEXITSTATUS(status);
 543 }
 544
 545 static int run_perf_stat(int argc __maybe_unused, const char **argv)
 546 {
 547         int ret;
 548
 549         if (pre_cmd) {
 550                 ret = system(pre_cmd);
 551                 if (ret)
 552                         return ret;
 553         }
 554
 555         if (sync_run)
 556                 sync();
 557
 558         ret = __run_perf_stat(argc, argv);
 559         if (ret)
 560                 return ret;
 561
 562         if (post_cmd) {
 563                 ret = system(post_cmd);
 564                 if (ret)
 565                         return ret;
 566         }
 567
 568         return ret;
 569 }
 570
 571 static void print_noise_pct(double total, double avg)
 572 {
 573         double pct = rel_stddev_stats(total, avg);
 574
 575         if (csv_output)
 576                 fprintf(output, "%s%.2f%%", csv_sep, pct);
 577         else if (pct)
 578                 fprintf(output, "  ( +-%6.2f%% )", pct);
 579 }
 580
 581 static void print_noise(struct perf_evsel *evsel, double avg)
 582 {
 583         struct perf_stat *ps;
 584
 585         if (run_count == 1)
 586                 return;
 587
 588         ps = evsel->priv;
 589         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 590 }
 591
 592 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 593 {
 594         switch (aggr_mode) {
 595         case AGGR_CORE:
 596                 fprintf(output, "S%d-C%*d%s%*d%s",
 597                         cpu_map__id_to_socket(id),
 598                         csv_output ? 0 : -8,
 599                         cpu_map__id_to_cpu(id),
 600                         csv_sep,
 601                         csv_output ? 0 : 4,
 602                         nr,
 603                         csv_sep);
 604                 break;
 605         case AGGR_SOCKET:
 606                 fprintf(output, "S%*d%s%*d%s",
 607                         csv_output ? 0 : -5,
 608                         id,
 609                         csv_sep,
 610                         csv_output ? 0 : 4,
 611                         nr,
 612                         csv_sep);
 613                         break;
 614         case AGGR_NONE:
 615                 fprintf(output, "CPU%*d%s",
 616                         csv_output ? 0 : -4,
 617                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 618                 break;
 619         case AGGR_GLOBAL:
 620         default:
 621                 break;
 622         }
 623 }
 624
 625 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 626 {
 627         double msecs = avg / 1e6;
 628         const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
 629
 630         aggr_printout(evsel, cpu, nr);
 631
 632         fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
 633
 634         if (evsel->cgrp)
 635                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 636
 637         if (csv_output || interval)
 638                 return;
 639
 640         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 641                 fprintf(output, " # %8.3f CPUs utilized          ",
 642                         avg / avg_stats(&walltime_nsecs_stats));
 643         else
 644                 fprintf(output, "                                   ");
 645 }
 646
 647 /* used for get_ratio_color() */
 648 enum grc_type {
 649         GRC_STALLED_CYCLES_FE,
 650         GRC_STALLED_CYCLES_BE,
 651         GRC_CACHE_MISSES,
 652         GRC_MAX_NR
 653 };
 654
 655 static const char *get_ratio_color(enum grc_type type, double ratio)
 656 {
 657         static const double grc_table[GRC_MAX_NR][3] = {
 658                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
 659                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
 660                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
 661         };
 662         const char *color = PERF_COLOR_NORMAL;
 663
 664         if (ratio > grc_table[type][0])
 665                 color = PERF_COLOR_RED;
 666         else if (ratio > grc_table[type][1])
 667                 color = PERF_COLOR_MAGENTA;
 668         else if (ratio > grc_table[type][2])
 669                 color = PERF_COLOR_YELLOW;
 670
 671         return color;
 672 }
 673
 674 static void print_stalled_cycles_frontend(int cpu,
 675                                           struct perf_evsel *evsel
 676                                           __maybe_unused, double avg)
 677 {
 678         double total, ratio = 0.0;
 679         const char *color;
 680
 681         total = avg_stats(&runtime_cycles_stats[cpu]);
 682
 683         if (total)
 684                 ratio = avg / total * 100.0;
 685
 686         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
 687
 688         fprintf(output, " #  ");
 689         color_fprintf(output, color, "%6.2f%%", ratio);
 690         fprintf(output, " frontend cycles idle   ");
 691 }
 692
 693 static void print_stalled_cycles_backend(int cpu,
 694                                          struct perf_evsel *evsel
 695                                          __maybe_unused, double avg)
 696 {
 697         double total, ratio = 0.0;
 698         const char *color;
 699
 700         total = avg_stats(&runtime_cycles_stats[cpu]);
 701
 702         if (total)
 703                 ratio = avg / total * 100.0;
 704
 705         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
 706
 707         fprintf(output, " #  ");
 708         color_fprintf(output, color, "%6.2f%%", ratio);
 709         fprintf(output, " backend  cycles idle   ");
 710 }
 711
 712 static void print_branch_misses(int cpu,
 713                                 struct perf_evsel *evsel __maybe_unused,
 714                                 double avg)
 715 {
 716         double total, ratio = 0.0;
 717         const char *color;
 718
 719         total = avg_stats(&runtime_branches_stats[cpu]);
 720
 721         if (total)
 722                 ratio = avg / total * 100.0;
 723
 724         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 725
 726         fprintf(output, " #  ");
 727         color_fprintf(output, color, "%6.2f%%", ratio);
 728         fprintf(output, " of all branches        ");
 729 }
 730
 731 static void print_l1_dcache_misses(int cpu,
 732                                    struct perf_evsel *evsel __maybe_unused,
 733                                    double avg)
 734 {
 735         double total, ratio = 0.0;
 736         const char *color;
 737
 738         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
 739
 740         if (total)
 741                 ratio = avg / total * 100.0;
 742
 743         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 744
 745         fprintf(output, " #  ");
 746         color_fprintf(output, color, "%6.2f%%", ratio);
 747         fprintf(output, " of all L1-dcache hits  ");
 748 }
 749
 750 static void print_l1_icache_misses(int cpu,
 751                                    struct perf_evsel *evsel __maybe_unused,
 752                                    double avg)
 753 {
 754         double total, ratio = 0.0;
 755         const char *color;
 756
 757         total = avg_stats(&runtime_l1_icache_stats[cpu]);
 758
 759         if (total)
 760                 ratio = avg / total * 100.0;
 761
 762         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 763
 764         fprintf(output, " #  ");
 765         color_fprintf(output, color, "%6.2f%%", ratio);
 766         fprintf(output, " of all L1-icache hits  ");
 767 }
 768
 769 static void print_dtlb_cache_misses(int cpu,
 770                                     struct perf_evsel *evsel __maybe_unused,
 771                                     double avg)
 772 {
 773         double total, ratio = 0.0;
 774         const char *color;
 775
 776         total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
 777
 778         if (total)
 779                 ratio = avg / total * 100.0;
 780
 781         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 782
 783         fprintf(output, " #  ");
 784         color_fprintf(output, color, "%6.2f%%", ratio);
 785         fprintf(output, " of all dTLB cache hits ");
 786 }
 787
 788 static void print_itlb_cache_misses(int cpu,
 789                                     struct perf_evsel *evsel __maybe_unused,
 790                                     double avg)
 791 {
 792         double total, ratio = 0.0;
 793         const char *color;
 794
 795         total = avg_stats(&runtime_itlb_cache_stats[cpu]);
 796
 797         if (total)
 798                 ratio = avg / total * 100.0;
 799
 800         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 801
 802         fprintf(output, " #  ");
 803         color_fprintf(output, color, "%6.2f%%", ratio);
 804         fprintf(output, " of all iTLB cache hits ");
 805 }
 806
 807 static void print_ll_cache_misses(int cpu,
 808                                   struct perf_evsel *evsel __maybe_unused,
 809                                   double avg)
 810 {
 811         double total, ratio = 0.0;
 812         const char *color;
 813
 814         total = avg_stats(&runtime_ll_cache_stats[cpu]);
 815
 816         if (total)
 817                 ratio = avg / total * 100.0;
 818
 819         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 820
 821         fprintf(output, " #  ");
 822         color_fprintf(output, color, "%6.2f%%", ratio);
 823         fprintf(output, " of all LL-cache hits   ");
 824 }
 825
 826 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 827 {
 828         double total, ratio = 0.0;
 829         const char *fmt;
 830
 831         if (csv_output)
 832                 fmt = "%.0f%s%s";
 833         else if (big_num)
 834                 fmt = "%'18.0f%s%-25s";
 835         else
 836                 fmt = "%18.0f%s%-25s";
 837
 838         aggr_printout(evsel, cpu, nr);
 839
 840         if (aggr_mode == AGGR_GLOBAL)
 841                 cpu = 0;
 842
 843         fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
 844
 845         if (evsel->cgrp)
 846                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 847
 848         if (csv_output || interval)
 849                 return;
 850
 851         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 852                 total = avg_stats(&runtime_cycles_stats[cpu]);
 853                 if (total)
 854                         ratio = avg / total;
 855
 856                 fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
 857
 858                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
 859                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
 860
 861                 if (total && avg) {
 862                         ratio = total / avg;
 863                         fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
 864                 }
 865
 866         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 867                         runtime_branches_stats[cpu].n != 0) {
 868                 print_branch_misses(cpu, evsel, avg);
 869         } else if (
 870                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 871                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
 872                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 873                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 874                         runtime_l1_dcache_stats[cpu].n != 0) {
 875                 print_l1_dcache_misses(cpu, evsel, avg);
 876         } else if (
 877                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 878                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
 879                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 880                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 881                         runtime_l1_icache_stats[cpu].n != 0) {
 882                 print_l1_icache_misses(cpu, evsel, avg);
 883         } else if (
 884                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 885                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
 886                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 887                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 888                         runtime_dtlb_cache_stats[cpu].n != 0) {
 889                 print_dtlb_cache_misses(cpu, evsel, avg);
 890         } else if (
 891                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 892                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
 893                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 894                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 895                         runtime_itlb_cache_stats[cpu].n != 0) {
 896                 print_itlb_cache_misses(cpu, evsel, avg);
 897         } else if (
 898                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 899                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
 900                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 901                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 902                         runtime_ll_cache_stats[cpu].n != 0) {
 903                 print_ll_cache_misses(cpu, evsel, avg);
 904         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
 905                         runtime_cacherefs_stats[cpu].n != 0) {
 906                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
 907
 908                 if (total)
 909                         ratio = avg * 100 / total;
 910
 911                 fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
 912
 913         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
 914                 print_stalled_cycles_frontend(cpu, evsel, avg);
 915         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
 916                 print_stalled_cycles_backend(cpu, evsel, avg);
 917         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
 918                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 919
 920                 if (total)
 921                         ratio = 1.0 * avg / total;
 922
 923                 fprintf(output, " # %8.3f GHz                    ", ratio);
 924         } else if (runtime_nsecs_stats[cpu].n != 0) {
 925                 char unit = 'M';
 926
 927                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 928
 929                 if (total)
 930                         ratio = 1000.0 * avg / total;
 931                 if (ratio < 0.001) {
 932                         ratio *= 1000;
 933                         unit = 'K';
 934                 }
 935
 936                 fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
 937         } else {
 938                 fprintf(output, "                                   ");
 939         }
 940 }
 941
 942 static void print_aggr(char *prefix)
 943 {
 944         struct perf_evsel *counter;
 945         int cpu, cpu2, s, s2, id, nr;
 946         u64 ena, run, val;
 947
 948         if (!(aggr_map || aggr_get_id))
 949                 return;
 950
 951         for (s = 0; s < aggr_map->nr; s++) {
 952                 id = aggr_map->map[s];
 953                 list_for_each_entry(counter, &evsel_list->entries, node) {
 954                         val = ena = run = 0;
 955                         nr = 0;
 956                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 957                                 cpu2 = perf_evsel__cpus(counter)->map[cpu];
 958                                 s2 = aggr_get_id(evsel_list->cpus, cpu2);
 959                                 if (s2 != id)
 960                                         continue;
 961                                 val += counter->counts->cpu[cpu].val;
 962                                 ena += counter->counts->cpu[cpu].ena;
 963                                 run += counter->counts->cpu[cpu].run;
 964                                 nr++;
 965                         }
 966                         if (prefix)
 967                                 fprintf(output, "%s", prefix);
 968
 969                         if (run == 0 || ena == 0) {
 970                                 aggr_printout(counter, id, nr);
 971
 972                                 fprintf(output, "%*s%s%*s",
 973                                         csv_output ? 0 : 18,
 974                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
 975                                         csv_sep,
 976                                         csv_output ? 0 : -24,
 977                                         perf_evsel__name(counter));
 978
 979                                 if (counter->cgrp)
 980                                         fprintf(output, "%s%s",
 981                                                 csv_sep, counter->cgrp->name);
 982
 983                                 fputc('\n', output);
 984                                 continue;
 985                         }
 986
 987                         if (nsec_counter(counter))
 988                                 nsec_printout(id, nr, counter, val);
 989                         else
 990                                 abs_printout(id, nr, counter, val);
 991
 992                         if (!csv_output) {
 993                                 print_noise(counter, 1.0);
 994
 995                                 if (run != ena)
 996                                         fprintf(output, "  (%.2f%%)",
 997                                                 100.0 * run / ena);
 998                         }
 999                         fputc('\n', output);
1000                 }
1001         }
1002 }
1003
1004 /*
1005  * Print out the results of a single counter:
1006  * aggregated counts in system-wide mode
1007  */
1008 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1009 {
1010         struct perf_stat *ps = counter->priv;
1011         double avg = avg_stats(&ps->res_stats[0]);
1012         int scaled = counter->counts->scaled;
1013
1014         if (prefix)
1015                 fprintf(output, "%s", prefix);
1016
1017         if (scaled == -1) {
1018                 fprintf(output, "%*s%s%*s",
1019                         csv_output ? 0 : 18,
1020                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1021                         csv_sep,
1022                         csv_output ? 0 : -24,
1023                         perf_evsel__name(counter));
1024
1025                 if (counter->cgrp)
1026                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1027
1028                 fputc('\n', output);
1029                 return;
1030         }
1031
1032         if (nsec_counter(counter))
1033                 nsec_printout(-1, 0, counter, avg);
1034         else
1035                 abs_printout(-1, 0, counter, avg);
1036
1037         print_noise(counter, avg);
1038
1039         if (csv_output) {
1040                 fputc('\n', output);
1041                 return;
1042         }
1043
1044         if (scaled) {
1045                 double avg_enabled, avg_running;
1046
1047                 avg_enabled = avg_stats(&ps->res_stats[1]);
1048                 avg_running = avg_stats(&ps->res_stats[2]);
1049
1050                 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1051         }
1052         fprintf(output, "\n");
1053 }
1054
1055 /*
1056  * Print out the results of a single counter:
1057  * does not use aggregated count in system-wide
1058  */
1059 static void print_counter(struct perf_evsel *counter, char *prefix)
1060 {
1061         u64 ena, run, val;
1062         int cpu;
1063
1064         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1065                 val = counter->counts->cpu[cpu].val;
1066                 ena = counter->counts->cpu[cpu].ena;
1067                 run = counter->counts->cpu[cpu].run;
1068
1069                 if (prefix)
1070                         fprintf(output, "%s", prefix);
1071
1072                 if (run == 0 || ena == 0) {
1073                         fprintf(output, "CPU%*d%s%*s%s%*s",
1074                                 csv_output ? 0 : -4,
1075                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1076                                 csv_output ? 0 : 18,
1077                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1078                                 csv_sep,
1079                                 csv_output ? 0 : -24,
1080                                 perf_evsel__name(counter));
1081
1082                         if (counter->cgrp)
1083                                 fprintf(output, "%s%s",
1084                                         csv_sep, counter->cgrp->name);
1085
1086                         fputc('\n', output);
1087                         continue;
1088                 }
1089
1090                 if (nsec_counter(counter))
1091                         nsec_printout(cpu, 0, counter, val);
1092                 else
1093                         abs_printout(cpu, 0, counter, val);
1094
1095                 if (!csv_output) {
1096                         print_noise(counter, 1.0);
1097
1098                         if (run != ena)
1099                                 fprintf(output, "  (%.2f%%)",
1100                                         100.0 * run / ena);
1101                 }
1102                 fputc('\n', output);
1103         }
1104 }
1105
1106 static void print_stat(int argc, const char **argv)
1107 {
1108         struct perf_evsel *counter;
1109         int i;
1110
1111         fflush(stdout);
1112
1113         if (!csv_output) {
1114                 fprintf(output, "\n");
1115                 fprintf(output, " Performance counter stats for ");
1116                 if (!perf_target__has_task(&target)) {
1117                         fprintf(output, "\'%s", argv[0]);
1118                         for (i = 1; i < argc; i++)
1119                                 fprintf(output, " %s", argv[i]);
1120                 } else if (target.pid)
1121                         fprintf(output, "process id \'%s", target.pid);
1122                 else
1123                         fprintf(output, "thread id \'%s", target.tid);
1124
1125                 fprintf(output, "\'");
1126                 if (run_count > 1)
1127                         fprintf(output, " (%d runs)", run_count);
1128                 fprintf(output, ":\n\n");
1129         }
1130
1131         switch (aggr_mode) {
1132         case AGGR_CORE:
1133         case AGGR_SOCKET:
1134                 print_aggr(NULL);
1135                 break;
1136         case AGGR_GLOBAL:
1137                 list_for_each_entry(counter, &evsel_list->entries, node)
1138                         print_counter_aggr(counter, NULL);
1139                 break;
1140         case AGGR_NONE:
1141                 list_for_each_entry(counter, &evsel_list->entries, node)
1142                         print_counter(counter, NULL);
1143                 break;
1144         default:
1145                 break;
1146         }
1147
1148         if (!csv_output) {
1149                 if (!null_run)
1150                         fprintf(output, "\n");
1151                 fprintf(output, " %17.9f seconds time elapsed",
1152                                 avg_stats(&walltime_nsecs_stats)/1e9);
1153                 if (run_count > 1) {
1154                         fprintf(output, "                                        ");
1155                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1156                                         avg_stats(&walltime_nsecs_stats));
1157                 }
1158                 fprintf(output, "\n\n");
1159         }
1160 }
1161
1162 static volatile int signr = -1;
1163
1164 static void skip_signal(int signo)
1165 {
1166         if ((child_pid == -1) || interval)
1167                 done = 1;
1168
1169         signr = signo;
1170         /*
1171          * render child_pid harmless
1172          * won't send SIGTERM to a random
1173          * process in case of race condition
1174          * and fast PID recycling
1175          */
1176         child_pid = -1;
1177 }
1178
1179 static void sig_atexit(void)
1180 {
1181         sigset_t set, oset;
1182
1183         /*
1184          * avoid race condition with SIGCHLD handler
1185          * in skip_signal() which is modifying child_pid
1186          * goal is to avoid send SIGTERM to a random
1187          * process
1188          */
1189         sigemptyset(&set);
1190         sigaddset(&set, SIGCHLD);
1191         sigprocmask(SIG_BLOCK, &set, &oset);
1192
1193         if (child_pid != -1)
1194                 kill(child_pid, SIGTERM);
1195
1196         sigprocmask(SIG_SETMASK, &oset, NULL);
1197
1198         if (signr == -1)
1199                 return;
1200
1201         signal(signr, SIG_DFL);
1202         kill(getpid(), signr);
1203 }
1204
1205 static int stat__set_big_num(const struct option *opt __maybe_unused,
1206                              const char *s __maybe_unused, int unset)
1207 {
1208         big_num_opt = unset ? 0 : 1;
1209         return 0;
1210 }
1211
1212 static int perf_stat_init_aggr_mode(void)
1213 {
1214         switch (aggr_mode) {
1215         case AGGR_SOCKET:
1216                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1217                         perror("cannot build socket map");
1218                         return -1;
1219                 }
1220                 aggr_get_id = cpu_map__get_socket;
1221                 break;
1222         case AGGR_CORE:
1223                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1224                         perror("cannot build core map");
1225                         return -1;
1226                 }
1227                 aggr_get_id = cpu_map__get_core;
1228                 break;
1229         case AGGR_NONE:
1230         case AGGR_GLOBAL:
1231         default:
1232                 break;
1233         }
1234         return 0;
1235 }
1236
1237
1238 /*
1239  * Add default attributes, if there were no attributes specified or
1240  * if -d/--detailed, -d -d or -d -d -d is used:
1241  */
1242 static int add_default_attributes(void)
1243 {
1244         struct perf_event_attr default_attrs[] = {
1245
1246   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1247   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1248   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1249   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1250
1251   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1252   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1253   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1254   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1255   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1256   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1257
1258 };
1259
1260 /*
1261  * Detailed stats (-d), covering the L1 and last level data caches:
1262  */
1263         struct perf_event_attr detailed_attrs[] = {
1264
1265   { .type = PERF_TYPE_HW_CACHE,
1266     .config =
1267          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1268         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1269         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1270
1271   { .type = PERF_TYPE_HW_CACHE,
1272     .config =
1273          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1274         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1275         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1276
1277   { .type = PERF_TYPE_HW_CACHE,
1278     .config =
1279          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1280         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1281         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1282
1283   { .type = PERF_TYPE_HW_CACHE,
1284     .config =
1285          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1286         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1287         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1288 };
1289
1290 /*
1291  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1292  */
1293         struct perf_event_attr very_detailed_attrs[] = {
1294
1295   { .type = PERF_TYPE_HW_CACHE,
1296     .config =
1297          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1298         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1299         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1300
1301   { .type = PERF_TYPE_HW_CACHE,
1302     .config =
1303          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1304         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1305         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1306
1307   { .type = PERF_TYPE_HW_CACHE,
1308     .config =
1309          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1310         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1311         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1312
1313   { .type = PERF_TYPE_HW_CACHE,
1314     .config =
1315          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1316         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1317         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1318
1319   { .type = PERF_TYPE_HW_CACHE,
1320     .config =
1321          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1322         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1323         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1324
1325   { .type = PERF_TYPE_HW_CACHE,
1326     .config =
1327          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1328         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1329         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1330
1331 };
1332
1333 /*
1334  * Very, very detailed stats (-d -d -d), adding prefetch events:
1335  */
1336         struct perf_event_attr very_very_detailed_attrs[] = {
1337
1338   { .type = PERF_TYPE_HW_CACHE,
1339     .config =
1340          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1341         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1342         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1343
1344   { .type = PERF_TYPE_HW_CACHE,
1345     .config =
1346          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1347         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1348         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1349 };
1350
1351         /* Set attrs if no event is selected and !null_run: */
1352         if (null_run)
1353                 return 0;
1354
1355         if (!evsel_list->nr_entries) {
1356                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1357                         return -1;
1358         }
1359
1360         /* Detailed events get appended to the event list: */
1361
1362         if (detailed_run <  1)
1363                 return 0;
1364
1365         /* Append detailed run extra attributes: */
1366         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1367                 return -1;
1368
1369         if (detailed_run < 2)
1370                 return 0;
1371
1372         /* Append very detailed run extra attributes: */
1373         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1374                 return -1;
1375
1376         if (detailed_run < 3)
1377                 return 0;
1378
1379         /* Append very, very detailed run extra attributes: */
1380         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1381 }
1382
1383 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1384 {
1385         bool append_file = false;
1386         int output_fd = 0;
1387         const char *output_name = NULL;
1388         const struct option options[] = {
1389         OPT_CALLBACK('e', "event", &evsel_list, "event",
1390                      "event selector. use 'perf list' to list available events",
1391                      parse_events_option),
1392         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1393                      "event filter", parse_filter),
1394         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1395                     "child tasks do not inherit counters"),
1396         OPT_STRING('p', "pid", &target.pid, "pid",
1397                    "stat events on existing process id"),
1398         OPT_STRING('t', "tid", &target.tid, "tid",
1399                    "stat events on existing thread id"),
1400         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1401                     "system-wide collection from all CPUs"),
1402         OPT_BOOLEAN('g', "group", &group,
1403                     "put the counters into a counter group"),
1404         OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1405         OPT_INCR('v', "verbose", &verbose,
1406                     "be more verbose (show counter open errors, etc)"),
1407         OPT_INTEGER('r', "repeat", &run_count,
1408                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1409         OPT_BOOLEAN('n', "null", &null_run,
1410                     "null run - dont start any counters"),
1411         OPT_INCR('d', "detailed", &detailed_run,
1412                     "detailed run - start a lot of events"),
1413         OPT_BOOLEAN('S', "sync", &sync_run,
1414                     "call sync() before starting a run"),
1415         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1416                            "print large numbers with thousands\' separators",
1417                            stat__set_big_num),
1418         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1419                     "list of cpus to monitor in system-wide"),
1420         OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1421                     "disable CPU count aggregation", AGGR_NONE),
1422         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1423                    "print counts with custom separator"),
1424         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1425                      "monitor event in cgroup name only", parse_cgroups),
1426         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1427         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1428         OPT_INTEGER(0, "log-fd", &output_fd,
1429                     "log output to fd, instead of stderr"),
1430         OPT_STRING(0, "pre", &pre_cmd, "command",
1431                         "command to run prior to the measured command"),
1432         OPT_STRING(0, "post", &post_cmd, "command",
1433                         "command to run after to the measured command"),
1434         OPT_UINTEGER('I', "interval-print", &interval,
1435                     "print counts at regular interval in ms (>= 100)"),
1436         OPT_SET_UINT(0, "per-socket", &aggr_mode,
1437                      "aggregate counts per processor socket", AGGR_SOCKET),
1438         OPT_SET_UINT(0, "per-core", &aggr_mode,
1439                      "aggregate counts per physical processor core", AGGR_CORE),
1440         OPT_UINTEGER('D', "delay", &initial_delay,
1441                      "ms to wait before starting measurement after program start"),
1442         OPT_END()
1443         };
1444         const char * const stat_usage[] = {
1445                 "perf stat [<options>] [<command>]",
1446                 NULL
1447         };
1448         int status = -ENOMEM, run_idx;
1449         const char *mode;
1450
1451         setlocale(LC_ALL, "");
1452
1453         evsel_list = perf_evlist__new();
1454         if (evsel_list == NULL)
1455                 return -ENOMEM;
1456
1457         argc = parse_options(argc, argv, options, stat_usage,
1458                 PARSE_OPT_STOP_AT_NON_OPTION);
1459
1460         output = stderr;
1461         if (output_name && strcmp(output_name, "-"))
1462                 output = NULL;
1463
1464         if (output_name && output_fd) {
1465                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1466                 usage_with_options(stat_usage, options);
1467         }
1468
1469         if (output_fd < 0) {
1470                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1471                 usage_with_options(stat_usage, options);
1472         }
1473
1474         if (!output) {
1475                 struct timespec tm;
1476                 mode = append_file ? "a" : "w";
1477
1478                 output = fopen(output_name, mode);
1479                 if (!output) {
1480                         perror("failed to create output file");
1481                         return -1;
1482                 }
1483                 clock_gettime(CLOCK_REALTIME, &tm);
1484                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1485         } else if (output_fd > 0) {
1486                 mode = append_file ? "a" : "w";
1487                 output = fdopen(output_fd, mode);
1488                 if (!output) {
1489                         perror("Failed opening logfd");
1490                         return -errno;
1491                 }
1492         }
1493
1494         if (csv_sep) {
1495                 csv_output = true;
1496                 if (!strcmp(csv_sep, "\\t"))
1497                         csv_sep = "\t";
1498         } else
1499                 csv_sep = DEFAULT_SEPARATOR;
1500
1501         /*
1502          * let the spreadsheet do the pretty-printing
1503          */
1504         if (csv_output) {
1505                 /* User explicitly passed -B? */
1506                 if (big_num_opt == 1) {
1507                         fprintf(stderr, "-B option not supported with -x\n");
1508                         usage_with_options(stat_usage, options);
1509                 } else /* Nope, so disable big number formatting */
1510                         big_num = false;
1511         } else if (big_num_opt == 0) /* User passed --no-big-num */
1512                 big_num = false;
1513
1514         if (!argc && !perf_target__has_task(&target))
1515                 usage_with_options(stat_usage, options);
1516         if (run_count < 0) {
1517                 usage_with_options(stat_usage, options);
1518         } else if (run_count == 0) {
1519                 forever = true;
1520                 run_count = 1;
1521         }
1522
1523         /* no_aggr, cgroup are for system-wide only */
1524         if ((aggr_mode != AGGR_GLOBAL || nr_cgroups)
1525              && !perf_target__has_cpu(&target)) {
1526                 fprintf(stderr, "both cgroup and no-aggregation "
1527                         "modes only available in system-wide mode\n");
1528
1529                 usage_with_options(stat_usage, options);
1530                 return -1;
1531         }
1532
1533         if (add_default_attributes())
1534                 goto out;
1535
1536         perf_target__validate(&target);
1537
1538         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1539                 if (perf_target__has_task(&target))
1540                         pr_err("Problems finding threads of monitor\n");
1541                 if (perf_target__has_cpu(&target))
1542                         perror("failed to parse CPUs map");
1543
1544                 usage_with_options(stat_usage, options);
1545                 return -1;
1546         }
1547         if (interval && interval < 100) {
1548                 pr_err("print interval must be >= 100ms\n");
1549                 usage_with_options(stat_usage, options);
1550                 return -1;
1551         }
1552
1553         if (perf_evlist__alloc_stats(evsel_list, interval))
1554                 goto out_free_maps;
1555
1556         if (perf_stat_init_aggr_mode())
1557                 goto out;
1558
1559         /*
1560          * We dont want to block the signals - that would cause
1561          * child tasks to inherit that and Ctrl-C would not work.
1562          * What we want is for Ctrl-C to work in the exec()-ed
1563          * task, but being ignored by perf stat itself:
1564          */
1565         atexit(sig_atexit);
1566         if (!forever)
1567                 signal(SIGINT,  skip_signal);
1568         signal(SIGCHLD, skip_signal);
1569         signal(SIGALRM, skip_signal);
1570         signal(SIGABRT, skip_signal);
1571
1572         status = 0;
1573         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1574                 if (run_count != 1 && verbose)
1575                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1576                                 run_idx + 1);
1577
1578                 status = run_perf_stat(argc, argv);
1579                 if (forever && status != -1) {
1580                         print_stat(argc, argv);
1581                         perf_stat__reset_stats(evsel_list);
1582                 }
1583         }
1584
1585         if (!forever && status != -1 && !interval)
1586                 print_stat(argc, argv);
1587
1588         perf_evlist__free_stats(evsel_list);
1589 out_free_maps:
1590         perf_evlist__delete_maps(evsel_list);
1591 out:
1592         perf_evlist__delete(evsel_list);
1593         return status;
1594 }