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/pmu.h"
  50 #include "util/event.h"
  51 #include "util/evlist.h"
  52 #include "util/evsel.h"
  53 #include "util/debug.h"
  54 #include "util/color.h"
  55 #include "util/stat.h"
  56 #include "util/header.h"
  57 #include "util/cpumap.h"
  58 #include "util/thread.h"
  59 #include "util/thread_map.h"
  60
  61 #include <stdlib.h>
  62 #include <sys/prctl.h>
  63 #include <locale.h>
  64
  65 #define DEFAULT_SEPARATOR       " "
  66 #define CNTR_NOT_SUPPORTED      "<not supported>"
  67 #define CNTR_NOT_COUNTED        "<not counted>"
  68
  69 static void print_stat(int argc, const char **argv);
  70 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
  71 static void print_counter(struct perf_evsel *counter, char *prefix);
  72 static void print_aggr(char *prefix);
  73
  74 /* Default events used for perf stat -T */
  75 static const char * const transaction_attrs[] = {
  76         "task-clock",
  77         "{"
  78         "instructions,"
  79         "cycles,"
  80         "cpu/cycles-t/,"
  81         "cpu/tx-start/,"
  82         "cpu/el-start/,"
  83         "cpu/cycles-ct/"
  84         "}"
  85 };
  86
  87 /* More limited version when the CPU does not have all events. */
  88 static const char * const transaction_limited_attrs[] = {
  89         "task-clock",
  90         "{"
  91         "instructions,"
  92         "cycles,"
  93         "cpu/cycles-t/,"
  94         "cpu/tx-start/"
  95         "}"
  96 };
  97
  98 /* must match transaction_attrs and the beginning limited_attrs */
  99 enum {
 100         T_TASK_CLOCK,
 101         T_INSTRUCTIONS,
 102         T_CYCLES,
 103         T_CYCLES_IN_TX,
 104         T_TRANSACTION_START,
 105         T_ELISION_START,
 106         T_CYCLES_IN_TX_CP,
 107 };
 108
 109 static struct perf_evlist       *evsel_list;
 110
 111 static struct perf_target       target = {
 112         .uid    = UINT_MAX,
 113 };
 114
 115 enum aggr_mode {
 116         AGGR_NONE,
 117         AGGR_GLOBAL,
 118         AGGR_SOCKET,
 119         AGGR_CORE,
 120 };
 121
 122 static int                      run_count                       =  1;
 123 static bool                     no_inherit                      = false;
 124 static bool                     scale                           =  true;
 125 static enum aggr_mode           aggr_mode                       = AGGR_GLOBAL;
 126 static volatile pid_t           child_pid                       = -1;
 127 static bool                     null_run                        =  false;
 128 static int                      detailed_run                    =  0;
 129 static bool                     transaction_run;
 130 static bool                     big_num                         =  true;
 131 static int                      big_num_opt                     =  -1;
 132 static const char               *csv_sep                        = NULL;
 133 static bool                     csv_output                      = false;
 134 static bool                     group                           = false;
 135 static FILE                     *output                         = NULL;
 136 static const char               *pre_cmd                        = NULL;
 137 static const char               *post_cmd                       = NULL;
 138 static bool                     sync_run                        = false;
 139 static unsigned int             interval                        = 0;
 140 static unsigned int             initial_delay                   = 0;
 141 static bool                     forever                         = false;
 142 static struct timespec          ref_time;
 143 static struct cpu_map           *aggr_map;
 144 static int                      (*aggr_get_id)(struct cpu_map *m, int cpu);
 145
 146 static volatile int done = 0;
 147
 148 struct perf_stat {
 149         struct stats      res_stats[3];
 150 };
 151
 152 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 153                                  struct timespec *b)
 154 {
 155         r->tv_sec = a->tv_sec - b->tv_sec;
 156         if (a->tv_nsec < b->tv_nsec) {
 157                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 158                 r->tv_sec--;
 159         } else {
 160                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 161         }
 162 }
 163
 164 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
 165 {
 166         return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
 167 }
 168
 169 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
 170 {
 171         return perf_evsel__cpus(evsel)->nr;
 172 }
 173
 174 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 175 {
 176         memset(evsel->priv, 0, sizeof(struct perf_stat));
 177 }
 178
 179 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 180 {
 181         evsel->priv = zalloc(sizeof(struct perf_stat));
 182         return evsel->priv == NULL ? -ENOMEM : 0;
 183 }
 184
 185 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 186 {
 187         free(evsel->priv);
 188         evsel->priv = NULL;
 189 }
 190
 191 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
 192 {
 193         void *addr;
 194         size_t sz;
 195
 196         sz = sizeof(*evsel->counts) +
 197              (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
 198
 199         addr = zalloc(sz);
 200         if (!addr)
 201                 return -ENOMEM;
 202
 203         evsel->prev_raw_counts =  addr;
 204
 205         return 0;
 206 }
 207
 208 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
 209 {
 210         free(evsel->prev_raw_counts);
 211         evsel->prev_raw_counts = NULL;
 212 }
 213
 214 static void perf_evlist__free_stats(struct perf_evlist *evlist)
 215 {
 216         struct perf_evsel *evsel;
 217
 218         list_for_each_entry(evsel, &evlist->entries, node) {
 219                 perf_evsel__free_stat_priv(evsel);
 220                 perf_evsel__free_counts(evsel);
 221                 perf_evsel__free_prev_raw_counts(evsel);
 222         }
 223 }
 224
 225 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
 226 {
 227         struct perf_evsel *evsel;
 228
 229         list_for_each_entry(evsel, &evlist->entries, node) {
 230                 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 231                     perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
 232                     (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
 233                         goto out_free;
 234         }
 235
 236         return 0;
 237
 238 out_free:
 239         perf_evlist__free_stats(evlist);
 240         return -1;
 241 }
 242
 243 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
 244 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
 245 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
 246 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
 247 static struct stats runtime_branches_stats[MAX_NR_CPUS];
 248 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
 249 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
 250 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
 251 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
 252 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
 253 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
 254 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
 255 static struct stats walltime_nsecs_stats;
 256 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
 257 static struct stats runtime_elision_stats[MAX_NR_CPUS];
 258
 259 static void perf_stat__reset_stats(struct perf_evlist *evlist)
 260 {
 261         struct perf_evsel *evsel;
 262
 263         list_for_each_entry(evsel, &evlist->entries, node) {
 264                 perf_evsel__reset_stat_priv(evsel);
 265                 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
 266         }
 267
 268         memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
 269         memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
 270         memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
 271         memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
 272         memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
 273         memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
 274         memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
 275         memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
 276         memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
 277         memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
 278         memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
 279         memset(runtime_cycles_in_tx_stats, 0,
 280                         sizeof(runtime_cycles_in_tx_stats));
 281         memset(runtime_transaction_stats, 0,
 282                 sizeof(runtime_transaction_stats));
 283         memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
 284         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
 285 }
 286
 287 static int create_perf_stat_counter(struct perf_evsel *evsel)
 288 {
 289         struct perf_event_attr *attr = &evsel->attr;
 290
 291         if (scale)
 292                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 293                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 294
 295         attr->inherit = !no_inherit;
 296
 297         if (perf_target__has_cpu(&target))
 298                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 299
 300         if (!perf_target__has_task(&target) &&
 301             perf_evsel__is_group_leader(evsel)) {
 302                 attr->disabled = 1;
 303                 if (!initial_delay)
 304                         attr->enable_on_exec = 1;
 305         }
 306
 307         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 308 }
 309
 310 /*
 311  * Does the counter have nsecs as a unit?
 312  */
 313 static inline int nsec_counter(struct perf_evsel *evsel)
 314 {
 315         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 316             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 317                 return 1;
 318
 319         return 0;
 320 }
 321
 322 static struct perf_evsel *nth_evsel(int n)
 323 {
 324         static struct perf_evsel **array;
 325         static int array_len;
 326         struct perf_evsel *ev;
 327         int j;
 328
 329         /* Assumes this only called when evsel_list does not change anymore. */
 330         if (!array) {
 331                 list_for_each_entry(ev, &evsel_list->entries, node)
 332                         array_len++;
 333                 array = malloc(array_len * sizeof(void *));
 334                 if (!array)
 335                         exit(ENOMEM);
 336                 j = 0;
 337                 list_for_each_entry(ev, &evsel_list->entries, node)
 338                         array[j++] = ev;
 339         }
 340         if (n < array_len)
 341                 return array[n];
 342         return NULL;
 343 }
 344
 345 /*
 346  * Update various tracking values we maintain to print
 347  * more semantic information such as miss/hit ratios,
 348  * instruction rates, etc:
 349  */
 350 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
 351 {
 352         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 353                 update_stats(&runtime_nsecs_stats[0], count[0]);
 354         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 355                 update_stats(&runtime_cycles_stats[0], count[0]);
 356         else if (transaction_run &&
 357                  perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
 358                 update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
 359         else if (transaction_run &&
 360                  perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
 361                 update_stats(&runtime_transaction_stats[0], count[0]);
 362         else if (transaction_run &&
 363                  perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
 364                 update_stats(&runtime_elision_stats[0], count[0]);
 365         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
 366                 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
 367         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 368                 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
 369         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 370                 update_stats(&runtime_branches_stats[0], count[0]);
 371         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 372                 update_stats(&runtime_cacherefs_stats[0], count[0]);
 373         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
 374                 update_stats(&runtime_l1_dcache_stats[0], count[0]);
 375         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
 376                 update_stats(&runtime_l1_icache_stats[0], count[0]);
 377         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
 378                 update_stats(&runtime_ll_cache_stats[0], count[0]);
 379         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
 380                 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
 381         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
 382                 update_stats(&runtime_itlb_cache_stats[0], count[0]);
 383 }
 384
 385 /*
 386  * Read out the results of a single counter:
 387  * aggregate counts across CPUs in system-wide mode
 388  */
 389 static int read_counter_aggr(struct perf_evsel *counter)
 390 {
 391         struct perf_stat *ps = counter->priv;
 392         u64 *count = counter->counts->aggr.values;
 393         int i;
 394
 395         if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
 396                                thread_map__nr(evsel_list->threads), scale) < 0)
 397                 return -1;
 398
 399         for (i = 0; i < 3; i++)
 400                 update_stats(&ps->res_stats[i], count[i]);
 401
 402         if (verbose) {
 403                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 404                         perf_evsel__name(counter), count[0], count[1], count[2]);
 405         }
 406
 407         /*
 408          * Save the full runtime - to allow normalization during printout:
 409          */
 410         update_shadow_stats(counter, count);
 411
 412         return 0;
 413 }
 414
 415 /*
 416  * Read out the results of a single counter:
 417  * do not aggregate counts across CPUs in system-wide mode
 418  */
 419 static int read_counter(struct perf_evsel *counter)
 420 {
 421         u64 *count;
 422         int cpu;
 423
 424         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 425                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 426                         return -1;
 427
 428                 count = counter->counts->cpu[cpu].values;
 429
 430                 update_shadow_stats(counter, count);
 431         }
 432
 433         return 0;
 434 }
 435
 436 static void print_interval(void)
 437 {
 438         static int num_print_interval;
 439         struct perf_evsel *counter;
 440         struct perf_stat *ps;
 441         struct timespec ts, rs;
 442         char prefix[64];
 443
 444         if (aggr_mode == AGGR_GLOBAL) {
 445                 list_for_each_entry(counter, &evsel_list->entries, node) {
 446                         ps = counter->priv;
 447                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
 448                         read_counter_aggr(counter);
 449                 }
 450         } else  {
 451                 list_for_each_entry(counter, &evsel_list->entries, node) {
 452                         ps = counter->priv;
 453                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
 454                         read_counter(counter);
 455                 }
 456         }
 457
 458         clock_gettime(CLOCK_MONOTONIC, &ts);
 459         diff_timespec(&rs, &ts, &ref_time);
 460         sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
 461
 462         if (num_print_interval == 0 && !csv_output) {
 463                 switch (aggr_mode) {
 464                 case AGGR_SOCKET:
 465                         fprintf(output, "#           time socket cpus             counts events\n");
 466                         break;
 467                 case AGGR_CORE:
 468                         fprintf(output, "#           time core         cpus             counts events\n");
 469                         break;
 470                 case AGGR_NONE:
 471                         fprintf(output, "#           time CPU                 counts events\n");
 472                         break;
 473                 case AGGR_GLOBAL:
 474                 default:
 475                         fprintf(output, "#           time             counts events\n");
 476                 }
 477         }
 478
 479         if (++num_print_interval == 25)
 480                 num_print_interval = 0;
 481
 482         switch (aggr_mode) {
 483         case AGGR_CORE:
 484         case AGGR_SOCKET:
 485                 print_aggr(prefix);
 486                 break;
 487         case AGGR_NONE:
 488                 list_for_each_entry(counter, &evsel_list->entries, node)
 489                         print_counter(counter, prefix);
 490                 break;
 491         case AGGR_GLOBAL:
 492         default:
 493                 list_for_each_entry(counter, &evsel_list->entries, node)
 494                         print_counter_aggr(counter, prefix);
 495         }
 496
 497         fflush(output);
 498 }
 499
 500 static void handle_initial_delay(void)
 501 {
 502         struct perf_evsel *counter;
 503
 504         if (initial_delay) {
 505                 const int ncpus = cpu_map__nr(evsel_list->cpus),
 506                         nthreads = thread_map__nr(evsel_list->threads);
 507
 508                 usleep(initial_delay * 1000);
 509                 list_for_each_entry(counter, &evsel_list->entries, node)
 510                         perf_evsel__enable(counter, ncpus, nthreads);
 511         }
 512 }
 513
 514 static int __run_perf_stat(int argc, const char **argv)
 515 {
 516         char msg[512];
 517         unsigned long long t0, t1;
 518         struct perf_evsel *counter;
 519         struct timespec ts;
 520         int status = 0;
 521         const bool forks = (argc > 0);
 522
 523         if (interval) {
 524                 ts.tv_sec  = interval / 1000;
 525                 ts.tv_nsec = (interval % 1000) * 1000000;
 526         } else {
 527                 ts.tv_sec  = 1;
 528                 ts.tv_nsec = 0;
 529         }
 530
 531         if (forks) {
 532                 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
 533                                                   false, false) < 0) {
 534                         perror("failed to prepare workload");
 535                         return -1;
 536                 }
 537                 child_pid = evsel_list->workload.pid;
 538         }
 539
 540         if (group)
 541                 perf_evlist__set_leader(evsel_list);
 542
 543         list_for_each_entry(counter, &evsel_list->entries, node) {
 544                 if (create_perf_stat_counter(counter) < 0) {
 545                         /*
 546                          * PPC returns ENXIO for HW counters until 2.6.37
 547                          * (behavior changed with commit b0a873e).
 548                          */
 549                         if (errno == EINVAL || errno == ENOSYS ||
 550                             errno == ENOENT || errno == EOPNOTSUPP ||
 551                             errno == ENXIO) {
 552                                 if (verbose)
 553                                         ui__warning("%s event is not supported by the kernel.\n",
 554                                                     perf_evsel__name(counter));
 555                                 counter->supported = false;
 556                                 continue;
 557                         }
 558
 559                         perf_evsel__open_strerror(counter, &target,
 560                                                   errno, msg, sizeof(msg));
 561                         ui__error("%s\n", msg);
 562
 563                         if (child_pid != -1)
 564                                 kill(child_pid, SIGTERM);
 565
 566                         return -1;
 567                 }
 568                 counter->supported = true;
 569         }
 570
 571         if (perf_evlist__apply_filters(evsel_list)) {
 572                 error("failed to set filter with %d (%s)\n", errno,
 573                         strerror(errno));
 574                 return -1;
 575         }
 576
 577         /*
 578          * Enable counters and exec the command:
 579          */
 580         t0 = rdclock();
 581         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 582
 583         if (forks) {
 584                 perf_evlist__start_workload(evsel_list);
 585                 handle_initial_delay();
 586
 587                 if (interval) {
 588                         while (!waitpid(child_pid, &status, WNOHANG)) {
 589                                 nanosleep(&ts, NULL);
 590                                 print_interval();
 591                         }
 592                 }
 593                 wait(&status);
 594                 if (WIFSIGNALED(status))
 595                         psignal(WTERMSIG(status), argv[0]);
 596         } else {
 597                 handle_initial_delay();
 598                 while (!done) {
 599                         nanosleep(&ts, NULL);
 600                         if (interval)
 601                                 print_interval();
 602                 }
 603         }
 604
 605         t1 = rdclock();
 606
 607         update_stats(&walltime_nsecs_stats, t1 - t0);
 608
 609         if (aggr_mode == AGGR_GLOBAL) {
 610                 list_for_each_entry(counter, &evsel_list->entries, node) {
 611                         read_counter_aggr(counter);
 612                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 613                                              thread_map__nr(evsel_list->threads));
 614                 }
 615         } else {
 616                 list_for_each_entry(counter, &evsel_list->entries, node) {
 617                         read_counter(counter);
 618                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
 619                 }
 620         }
 621
 622         return WEXITSTATUS(status);
 623 }
 624
 625 static int run_perf_stat(int argc __maybe_unused, const char **argv)
 626 {
 627         int ret;
 628
 629         if (pre_cmd) {
 630                 ret = system(pre_cmd);
 631                 if (ret)
 632                         return ret;
 633         }
 634
 635         if (sync_run)
 636                 sync();
 637
 638         ret = __run_perf_stat(argc, argv);
 639         if (ret)
 640                 return ret;
 641
 642         if (post_cmd) {
 643                 ret = system(post_cmd);
 644                 if (ret)
 645                         return ret;
 646         }
 647
 648         return ret;
 649 }
 650
 651 static void print_noise_pct(double total, double avg)
 652 {
 653         double pct = rel_stddev_stats(total, avg);
 654
 655         if (csv_output)
 656                 fprintf(output, "%s%.2f%%", csv_sep, pct);
 657         else if (pct)
 658                 fprintf(output, "  ( +-%6.2f%% )", pct);
 659 }
 660
 661 static void print_noise(struct perf_evsel *evsel, double avg)
 662 {
 663         struct perf_stat *ps;
 664
 665         if (run_count == 1)
 666                 return;
 667
 668         ps = evsel->priv;
 669         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 670 }
 671
 672 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 673 {
 674         switch (aggr_mode) {
 675         case AGGR_CORE:
 676                 fprintf(output, "S%d-C%*d%s%*d%s",
 677                         cpu_map__id_to_socket(id),
 678                         csv_output ? 0 : -8,
 679                         cpu_map__id_to_cpu(id),
 680                         csv_sep,
 681                         csv_output ? 0 : 4,
 682                         nr,
 683                         csv_sep);
 684                 break;
 685         case AGGR_SOCKET:
 686                 fprintf(output, "S%*d%s%*d%s",
 687                         csv_output ? 0 : -5,
 688                         id,
 689                         csv_sep,
 690                         csv_output ? 0 : 4,
 691                         nr,
 692                         csv_sep);
 693                         break;
 694         case AGGR_NONE:
 695                 fprintf(output, "CPU%*d%s",
 696                         csv_output ? 0 : -4,
 697                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 698                 break;
 699         case AGGR_GLOBAL:
 700         default:
 701                 break;
 702         }
 703 }
 704
 705 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 706 {
 707         double msecs = avg / 1e6;
 708         const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
 709
 710         aggr_printout(evsel, cpu, nr);
 711
 712         fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
 713
 714         if (evsel->cgrp)
 715                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 716
 717         if (csv_output || interval)
 718                 return;
 719
 720         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 721                 fprintf(output, " # %8.3f CPUs utilized          ",
 722                         avg / avg_stats(&walltime_nsecs_stats));
 723         else
 724                 fprintf(output, "                                   ");
 725 }
 726
 727 /* used for get_ratio_color() */
 728 enum grc_type {
 729         GRC_STALLED_CYCLES_FE,
 730         GRC_STALLED_CYCLES_BE,
 731         GRC_CACHE_MISSES,
 732         GRC_MAX_NR
 733 };
 734
 735 static const char *get_ratio_color(enum grc_type type, double ratio)
 736 {
 737         static const double grc_table[GRC_MAX_NR][3] = {
 738                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
 739                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
 740                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
 741         };
 742         const char *color = PERF_COLOR_NORMAL;
 743
 744         if (ratio > grc_table[type][0])
 745                 color = PERF_COLOR_RED;
 746         else if (ratio > grc_table[type][1])
 747                 color = PERF_COLOR_MAGENTA;
 748         else if (ratio > grc_table[type][2])
 749                 color = PERF_COLOR_YELLOW;
 750
 751         return color;
 752 }
 753
 754 static void print_stalled_cycles_frontend(int cpu,
 755                                           struct perf_evsel *evsel
 756                                           __maybe_unused, double avg)
 757 {
 758         double total, ratio = 0.0;
 759         const char *color;
 760
 761         total = avg_stats(&runtime_cycles_stats[cpu]);
 762
 763         if (total)
 764                 ratio = avg / total * 100.0;
 765
 766         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
 767
 768         fprintf(output, " #  ");
 769         color_fprintf(output, color, "%6.2f%%", ratio);
 770         fprintf(output, " frontend cycles idle   ");
 771 }
 772
 773 static void print_stalled_cycles_backend(int cpu,
 774                                          struct perf_evsel *evsel
 775                                          __maybe_unused, double avg)
 776 {
 777         double total, ratio = 0.0;
 778         const char *color;
 779
 780         total = avg_stats(&runtime_cycles_stats[cpu]);
 781
 782         if (total)
 783                 ratio = avg / total * 100.0;
 784
 785         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
 786
 787         fprintf(output, " #  ");
 788         color_fprintf(output, color, "%6.2f%%", ratio);
 789         fprintf(output, " backend  cycles idle   ");
 790 }
 791
 792 static void print_branch_misses(int cpu,
 793                                 struct perf_evsel *evsel __maybe_unused,
 794                                 double avg)
 795 {
 796         double total, ratio = 0.0;
 797         const char *color;
 798
 799         total = avg_stats(&runtime_branches_stats[cpu]);
 800
 801         if (total)
 802                 ratio = avg / total * 100.0;
 803
 804         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 805
 806         fprintf(output, " #  ");
 807         color_fprintf(output, color, "%6.2f%%", ratio);
 808         fprintf(output, " of all branches        ");
 809 }
 810
 811 static void print_l1_dcache_misses(int cpu,
 812                                    struct perf_evsel *evsel __maybe_unused,
 813                                    double avg)
 814 {
 815         double total, ratio = 0.0;
 816         const char *color;
 817
 818         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
 819
 820         if (total)
 821                 ratio = avg / total * 100.0;
 822
 823         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 824
 825         fprintf(output, " #  ");
 826         color_fprintf(output, color, "%6.2f%%", ratio);
 827         fprintf(output, " of all L1-dcache hits  ");
 828 }
 829
 830 static void print_l1_icache_misses(int cpu,
 831                                    struct perf_evsel *evsel __maybe_unused,
 832                                    double avg)
 833 {
 834         double total, ratio = 0.0;
 835         const char *color;
 836
 837         total = avg_stats(&runtime_l1_icache_stats[cpu]);
 838
 839         if (total)
 840                 ratio = avg / total * 100.0;
 841
 842         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 843
 844         fprintf(output, " #  ");
 845         color_fprintf(output, color, "%6.2f%%", ratio);
 846         fprintf(output, " of all L1-icache hits  ");
 847 }
 848
 849 static void print_dtlb_cache_misses(int cpu,
 850                                     struct perf_evsel *evsel __maybe_unused,
 851                                     double avg)
 852 {
 853         double total, ratio = 0.0;
 854         const char *color;
 855
 856         total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
 857
 858         if (total)
 859                 ratio = avg / total * 100.0;
 860
 861         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 862
 863         fprintf(output, " #  ");
 864         color_fprintf(output, color, "%6.2f%%", ratio);
 865         fprintf(output, " of all dTLB cache hits ");
 866 }
 867
 868 static void print_itlb_cache_misses(int cpu,
 869                                     struct perf_evsel *evsel __maybe_unused,
 870                                     double avg)
 871 {
 872         double total, ratio = 0.0;
 873         const char *color;
 874
 875         total = avg_stats(&runtime_itlb_cache_stats[cpu]);
 876
 877         if (total)
 878                 ratio = avg / total * 100.0;
 879
 880         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 881
 882         fprintf(output, " #  ");
 883         color_fprintf(output, color, "%6.2f%%", ratio);
 884         fprintf(output, " of all iTLB cache hits ");
 885 }
 886
 887 static void print_ll_cache_misses(int cpu,
 888                                   struct perf_evsel *evsel __maybe_unused,
 889                                   double avg)
 890 {
 891         double total, ratio = 0.0;
 892         const char *color;
 893
 894         total = avg_stats(&runtime_ll_cache_stats[cpu]);
 895
 896         if (total)
 897                 ratio = avg / total * 100.0;
 898
 899         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 900
 901         fprintf(output, " #  ");
 902         color_fprintf(output, color, "%6.2f%%", ratio);
 903         fprintf(output, " of all LL-cache hits   ");
 904 }
 905
 906 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 907 {
 908         double total, ratio = 0.0, total2;
 909         const char *fmt;
 910
 911         if (csv_output)
 912                 fmt = "%.0f%s%s";
 913         else if (big_num)
 914                 fmt = "%'18.0f%s%-25s";
 915         else
 916                 fmt = "%18.0f%s%-25s";
 917
 918         aggr_printout(evsel, cpu, nr);
 919
 920         if (aggr_mode == AGGR_GLOBAL)
 921                 cpu = 0;
 922
 923         fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
 924
 925         if (evsel->cgrp)
 926                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 927
 928         if (csv_output || interval)
 929                 return;
 930
 931         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 932                 total = avg_stats(&runtime_cycles_stats[cpu]);
 933                 if (total) {
 934                         ratio = avg / total;
 935                         fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
 936                 }
 937                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
 938                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
 939
 940                 if (total && avg) {
 941                         ratio = total / avg;
 942                         fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
 943                 }
 944
 945         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 946                         runtime_branches_stats[cpu].n != 0) {
 947                 print_branch_misses(cpu, evsel, avg);
 948         } else if (
 949                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 950                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
 951                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 952                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 953                         runtime_l1_dcache_stats[cpu].n != 0) {
 954                 print_l1_dcache_misses(cpu, evsel, avg);
 955         } else if (
 956                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 957                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
 958                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 959                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 960                         runtime_l1_icache_stats[cpu].n != 0) {
 961                 print_l1_icache_misses(cpu, evsel, avg);
 962         } else if (
 963                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 964                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
 965                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 966                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 967                         runtime_dtlb_cache_stats[cpu].n != 0) {
 968                 print_dtlb_cache_misses(cpu, evsel, avg);
 969         } else if (
 970                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 971                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
 972                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 973                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 974                         runtime_itlb_cache_stats[cpu].n != 0) {
 975                 print_itlb_cache_misses(cpu, evsel, avg);
 976         } else if (
 977                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 978                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
 979                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 980                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 981                         runtime_ll_cache_stats[cpu].n != 0) {
 982                 print_ll_cache_misses(cpu, evsel, avg);
 983         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
 984                         runtime_cacherefs_stats[cpu].n != 0) {
 985                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
 986
 987                 if (total)
 988                         ratio = avg * 100 / total;
 989
 990                 fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
 991
 992         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
 993                 print_stalled_cycles_frontend(cpu, evsel, avg);
 994         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
 995                 print_stalled_cycles_backend(cpu, evsel, avg);
 996         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
 997                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 998
 999                 if (total) {
1000                         ratio = avg / total;
1001                         fprintf(output, " # %8.3f GHz                    ", ratio);
1002                 }
1003         } else if (transaction_run &&
1004                    perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1005                 total = avg_stats(&runtime_cycles_stats[cpu]);
1006                 if (total)
1007                         fprintf(output,
1008                                 " #   %5.2f%% transactional cycles   ",
1009                                 100.0 * (avg / total));
1010         } else if (transaction_run &&
1011                    perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1012                 total = avg_stats(&runtime_cycles_stats[cpu]);
1013                 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1014                 if (total2 < avg)
1015                         total2 = avg;
1016                 if (total)
1017                         fprintf(output,
1018                                 " #   %5.2f%% aborted cycles         ",
1019                                 100.0 * ((total2-avg) / total));
1020         } else if (transaction_run &&
1021                    perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1022                    avg > 0 &&
1023                    runtime_cycles_in_tx_stats[cpu].n != 0) {
1024                 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1025
1026                 if (total)
1027                         ratio = total / avg;
1028
1029                 fprintf(output, " # %8.0f cycles / transaction   ", ratio);
1030         } else if (transaction_run &&
1031                    perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1032                    avg > 0 &&
1033                    runtime_cycles_in_tx_stats[cpu].n != 0) {
1034                 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1035
1036                 if (total)
1037                         ratio = total / avg;
1038
1039                 fprintf(output, " # %8.0f cycles / elision       ", ratio);
1040         } else if (runtime_nsecs_stats[cpu].n != 0) {
1041                 char unit = 'M';
1042
1043                 total = avg_stats(&runtime_nsecs_stats[cpu]);
1044
1045                 if (total)
1046                         ratio = 1000.0 * avg / total;
1047                 if (ratio < 0.001) {
1048                         ratio *= 1000;
1049                         unit = 'K';
1050                 }
1051
1052                 fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
1053         } else {
1054                 fprintf(output, "                                   ");
1055         }
1056 }
1057
1058 static void print_aggr(char *prefix)
1059 {
1060         struct perf_evsel *counter;
1061         int cpu, cpu2, s, s2, id, nr;
1062         u64 ena, run, val;
1063
1064         if (!(aggr_map || aggr_get_id))
1065                 return;
1066
1067         for (s = 0; s < aggr_map->nr; s++) {
1068                 id = aggr_map->map[s];
1069                 list_for_each_entry(counter, &evsel_list->entries, node) {
1070                         val = ena = run = 0;
1071                         nr = 0;
1072                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1073                                 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1074                                 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1075                                 if (s2 != id)
1076                                         continue;
1077                                 val += counter->counts->cpu[cpu].val;
1078                                 ena += counter->counts->cpu[cpu].ena;
1079                                 run += counter->counts->cpu[cpu].run;
1080                                 nr++;
1081                         }
1082                         if (prefix)
1083                                 fprintf(output, "%s", prefix);
1084
1085                         if (run == 0 || ena == 0) {
1086                                 aggr_printout(counter, id, nr);
1087
1088                                 fprintf(output, "%*s%s%*s",
1089                                         csv_output ? 0 : 18,
1090                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1091                                         csv_sep,
1092                                         csv_output ? 0 : -24,
1093                                         perf_evsel__name(counter));
1094
1095                                 if (counter->cgrp)
1096                                         fprintf(output, "%s%s",
1097                                                 csv_sep, counter->cgrp->name);
1098
1099                                 fputc('\n', output);
1100                                 continue;
1101                         }
1102
1103                         if (nsec_counter(counter))
1104                                 nsec_printout(id, nr, counter, val);
1105                         else
1106                                 abs_printout(id, nr, counter, val);
1107
1108                         if (!csv_output) {
1109                                 print_noise(counter, 1.0);
1110
1111                                 if (run != ena)
1112                                         fprintf(output, "  (%.2f%%)",
1113                                                 100.0 * run / ena);
1114                         }
1115                         fputc('\n', output);
1116                 }
1117         }
1118 }
1119
1120 /*
1121  * Print out the results of a single counter:
1122  * aggregated counts in system-wide mode
1123  */
1124 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1125 {
1126         struct perf_stat *ps = counter->priv;
1127         double avg = avg_stats(&ps->res_stats[0]);
1128         int scaled = counter->counts->scaled;
1129
1130         if (prefix)
1131                 fprintf(output, "%s", prefix);
1132
1133         if (scaled == -1) {
1134                 fprintf(output, "%*s%s%*s",
1135                         csv_output ? 0 : 18,
1136                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1137                         csv_sep,
1138                         csv_output ? 0 : -24,
1139                         perf_evsel__name(counter));
1140
1141                 if (counter->cgrp)
1142                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1143
1144                 fputc('\n', output);
1145                 return;
1146         }
1147
1148         if (nsec_counter(counter))
1149                 nsec_printout(-1, 0, counter, avg);
1150         else
1151                 abs_printout(-1, 0, counter, avg);
1152
1153         print_noise(counter, avg);
1154
1155         if (csv_output) {
1156                 fputc('\n', output);
1157                 return;
1158         }
1159
1160         if (scaled) {
1161                 double avg_enabled, avg_running;
1162
1163                 avg_enabled = avg_stats(&ps->res_stats[1]);
1164                 avg_running = avg_stats(&ps->res_stats[2]);
1165
1166                 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1167         }
1168         fprintf(output, "\n");
1169 }
1170
1171 /*
1172  * Print out the results of a single counter:
1173  * does not use aggregated count in system-wide
1174  */
1175 static void print_counter(struct perf_evsel *counter, char *prefix)
1176 {
1177         u64 ena, run, val;
1178         int cpu;
1179
1180         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1181                 val = counter->counts->cpu[cpu].val;
1182                 ena = counter->counts->cpu[cpu].ena;
1183                 run = counter->counts->cpu[cpu].run;
1184
1185                 if (prefix)
1186                         fprintf(output, "%s", prefix);
1187
1188                 if (run == 0 || ena == 0) {
1189                         fprintf(output, "CPU%*d%s%*s%s%*s",
1190                                 csv_output ? 0 : -4,
1191                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1192                                 csv_output ? 0 : 18,
1193                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1194                                 csv_sep,
1195                                 csv_output ? 0 : -24,
1196                                 perf_evsel__name(counter));
1197
1198                         if (counter->cgrp)
1199                                 fprintf(output, "%s%s",
1200                                         csv_sep, counter->cgrp->name);
1201
1202                         fputc('\n', output);
1203                         continue;
1204                 }
1205
1206                 if (nsec_counter(counter))
1207                         nsec_printout(cpu, 0, counter, val);
1208                 else
1209                         abs_printout(cpu, 0, counter, val);
1210
1211                 if (!csv_output) {
1212                         print_noise(counter, 1.0);
1213
1214                         if (run != ena)
1215                                 fprintf(output, "  (%.2f%%)",
1216                                         100.0 * run / ena);
1217                 }
1218                 fputc('\n', output);
1219         }
1220 }
1221
1222 static void print_stat(int argc, const char **argv)
1223 {
1224         struct perf_evsel *counter;
1225         int i;
1226
1227         fflush(stdout);
1228
1229         if (!csv_output) {
1230                 fprintf(output, "\n");
1231                 fprintf(output, " Performance counter stats for ");
1232                 if (target.system_wide)
1233                         fprintf(output, "\'system wide");
1234                 else if (target.cpu_list)
1235                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1236                 else if (!perf_target__has_task(&target)) {
1237                         fprintf(output, "\'%s", argv[0]);
1238                         for (i = 1; i < argc; i++)
1239                                 fprintf(output, " %s", argv[i]);
1240                 } else if (target.pid)
1241                         fprintf(output, "process id \'%s", target.pid);
1242                 else
1243                         fprintf(output, "thread id \'%s", target.tid);
1244
1245                 fprintf(output, "\'");
1246                 if (run_count > 1)
1247                         fprintf(output, " (%d runs)", run_count);
1248                 fprintf(output, ":\n\n");
1249         }
1250
1251         switch (aggr_mode) {
1252         case AGGR_CORE:
1253         case AGGR_SOCKET:
1254                 print_aggr(NULL);
1255                 break;
1256         case AGGR_GLOBAL:
1257                 list_for_each_entry(counter, &evsel_list->entries, node)
1258                         print_counter_aggr(counter, NULL);
1259                 break;
1260         case AGGR_NONE:
1261                 list_for_each_entry(counter, &evsel_list->entries, node)
1262                         print_counter(counter, NULL);
1263                 break;
1264         default:
1265                 break;
1266         }
1267
1268         if (!csv_output) {
1269                 if (!null_run)
1270                         fprintf(output, "\n");
1271                 fprintf(output, " %17.9f seconds time elapsed",
1272                                 avg_stats(&walltime_nsecs_stats)/1e9);
1273                 if (run_count > 1) {
1274                         fprintf(output, "                                        ");
1275                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1276                                         avg_stats(&walltime_nsecs_stats));
1277                 }
1278                 fprintf(output, "\n\n");
1279         }
1280 }
1281
1282 static volatile int signr = -1;
1283
1284 static void skip_signal(int signo)
1285 {
1286         if ((child_pid == -1) || interval)
1287                 done = 1;
1288
1289         signr = signo;
1290         /*
1291          * render child_pid harmless
1292          * won't send SIGTERM to a random
1293          * process in case of race condition
1294          * and fast PID recycling
1295          */
1296         child_pid = -1;
1297 }
1298
1299 static void sig_atexit(void)
1300 {
1301         sigset_t set, oset;
1302
1303         /*
1304          * avoid race condition with SIGCHLD handler
1305          * in skip_signal() which is modifying child_pid
1306          * goal is to avoid send SIGTERM to a random
1307          * process
1308          */
1309         sigemptyset(&set);
1310         sigaddset(&set, SIGCHLD);
1311         sigprocmask(SIG_BLOCK, &set, &oset);
1312
1313         if (child_pid != -1)
1314                 kill(child_pid, SIGTERM);
1315
1316         sigprocmask(SIG_SETMASK, &oset, NULL);
1317
1318         if (signr == -1)
1319                 return;
1320
1321         signal(signr, SIG_DFL);
1322         kill(getpid(), signr);
1323 }
1324
1325 static int stat__set_big_num(const struct option *opt __maybe_unused,
1326                              const char *s __maybe_unused, int unset)
1327 {
1328         big_num_opt = unset ? 0 : 1;
1329         return 0;
1330 }
1331
1332 static int perf_stat_init_aggr_mode(void)
1333 {
1334         switch (aggr_mode) {
1335         case AGGR_SOCKET:
1336                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1337                         perror("cannot build socket map");
1338                         return -1;
1339                 }
1340                 aggr_get_id = cpu_map__get_socket;
1341                 break;
1342         case AGGR_CORE:
1343                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1344                         perror("cannot build core map");
1345                         return -1;
1346                 }
1347                 aggr_get_id = cpu_map__get_core;
1348                 break;
1349         case AGGR_NONE:
1350         case AGGR_GLOBAL:
1351         default:
1352                 break;
1353         }
1354         return 0;
1355 }
1356
1357 static int setup_events(const char * const *attrs, unsigned len)
1358 {
1359         unsigned i;
1360
1361         for (i = 0; i < len; i++) {
1362                 if (parse_events(evsel_list, attrs[i]))
1363                         return -1;
1364         }
1365         return 0;
1366 }
1367
1368 /*
1369  * Add default attributes, if there were no attributes specified or
1370  * if -d/--detailed, -d -d or -d -d -d is used:
1371  */
1372 static int add_default_attributes(void)
1373 {
1374         struct perf_event_attr default_attrs[] = {
1375
1376   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1377   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1378   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1379   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1380
1381   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1382   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1383   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1384   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1385   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1386   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1387
1388 };
1389
1390 /*
1391  * Detailed stats (-d), covering the L1 and last level data caches:
1392  */
1393         struct perf_event_attr detailed_attrs[] = {
1394
1395   { .type = PERF_TYPE_HW_CACHE,
1396     .config =
1397          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1398         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1399         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1400
1401   { .type = PERF_TYPE_HW_CACHE,
1402     .config =
1403          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1404         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1405         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1406
1407   { .type = PERF_TYPE_HW_CACHE,
1408     .config =
1409          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1410         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1411         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1412
1413   { .type = PERF_TYPE_HW_CACHE,
1414     .config =
1415          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1416         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1417         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1418 };
1419
1420 /*
1421  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1422  */
1423         struct perf_event_attr very_detailed_attrs[] = {
1424
1425   { .type = PERF_TYPE_HW_CACHE,
1426     .config =
1427          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1428         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1429         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1430
1431   { .type = PERF_TYPE_HW_CACHE,
1432     .config =
1433          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1434         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1435         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1436
1437   { .type = PERF_TYPE_HW_CACHE,
1438     .config =
1439          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1440         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1441         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1442
1443   { .type = PERF_TYPE_HW_CACHE,
1444     .config =
1445          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1446         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1447         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1448
1449   { .type = PERF_TYPE_HW_CACHE,
1450     .config =
1451          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1452         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1453         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1454
1455   { .type = PERF_TYPE_HW_CACHE,
1456     .config =
1457          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1458         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1459         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1460
1461 };
1462
1463 /*
1464  * Very, very detailed stats (-d -d -d), adding prefetch events:
1465  */
1466         struct perf_event_attr very_very_detailed_attrs[] = {
1467
1468   { .type = PERF_TYPE_HW_CACHE,
1469     .config =
1470          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1471         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1472         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1473
1474   { .type = PERF_TYPE_HW_CACHE,
1475     .config =
1476          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1477         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1478         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1479 };
1480
1481         /* Set attrs if no event is selected and !null_run: */
1482         if (null_run)
1483                 return 0;
1484
1485         if (transaction_run) {
1486                 int err;
1487                 if (pmu_have_event("cpu", "cycles-ct") &&
1488                     pmu_have_event("cpu", "el-start"))
1489                         err = setup_events(transaction_attrs,
1490                                         ARRAY_SIZE(transaction_attrs));
1491                 else
1492                         err = setup_events(transaction_limited_attrs,
1493                                  ARRAY_SIZE(transaction_limited_attrs));
1494                 if (err < 0) {
1495                         fprintf(stderr, "Cannot set up transaction events\n");
1496                         return -1;
1497                 }
1498                 return 0;
1499         }
1500
1501         if (!evsel_list->nr_entries) {
1502                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1503                         return -1;
1504         }
1505
1506         /* Detailed events get appended to the event list: */
1507
1508         if (detailed_run <  1)
1509                 return 0;
1510
1511         /* Append detailed run extra attributes: */
1512         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1513                 return -1;
1514
1515         if (detailed_run < 2)
1516                 return 0;
1517
1518         /* Append very detailed run extra attributes: */
1519         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1520                 return -1;
1521
1522         if (detailed_run < 3)
1523                 return 0;
1524
1525         /* Append very, very detailed run extra attributes: */
1526         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1527 }
1528
1529 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1530 {
1531         bool append_file = false;
1532         int output_fd = 0;
1533         const char *output_name = NULL;
1534         const struct option options[] = {
1535         OPT_BOOLEAN('T', "transaction", &transaction_run,
1536                     "hardware transaction statistics"),
1537         OPT_CALLBACK('e', "event", &evsel_list, "event",
1538                      "event selector. use 'perf list' to list available events",
1539                      parse_events_option),
1540         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1541                      "event filter", parse_filter),
1542         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1543                     "child tasks do not inherit counters"),
1544         OPT_STRING('p', "pid", &target.pid, "pid",
1545                    "stat events on existing process id"),
1546         OPT_STRING('t', "tid", &target.tid, "tid",
1547                    "stat events on existing thread id"),
1548         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1549                     "system-wide collection from all CPUs"),
1550         OPT_BOOLEAN('g', "group", &group,
1551                     "put the counters into a counter group"),
1552         OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1553         OPT_INCR('v', "verbose", &verbose,
1554                     "be more verbose (show counter open errors, etc)"),
1555         OPT_INTEGER('r', "repeat", &run_count,
1556                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1557         OPT_BOOLEAN('n', "null", &null_run,
1558                     "null run - dont start any counters"),
1559         OPT_INCR('d', "detailed", &detailed_run,
1560                     "detailed run - start a lot of events"),
1561         OPT_BOOLEAN('S', "sync", &sync_run,
1562                     "call sync() before starting a run"),
1563         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1564                            "print large numbers with thousands\' separators",
1565                            stat__set_big_num),
1566         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1567                     "list of cpus to monitor in system-wide"),
1568         OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1569                     "disable CPU count aggregation", AGGR_NONE),
1570         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1571                    "print counts with custom separator"),
1572         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1573                      "monitor event in cgroup name only", parse_cgroups),
1574         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1575         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1576         OPT_INTEGER(0, "log-fd", &output_fd,
1577                     "log output to fd, instead of stderr"),
1578         OPT_STRING(0, "pre", &pre_cmd, "command",
1579                         "command to run prior to the measured command"),
1580         OPT_STRING(0, "post", &post_cmd, "command",
1581                         "command to run after to the measured command"),
1582         OPT_UINTEGER('I', "interval-print", &interval,
1583                     "print counts at regular interval in ms (>= 100)"),
1584         OPT_SET_UINT(0, "per-socket", &aggr_mode,
1585                      "aggregate counts per processor socket", AGGR_SOCKET),
1586         OPT_SET_UINT(0, "per-core", &aggr_mode,
1587                      "aggregate counts per physical processor core", AGGR_CORE),
1588         OPT_UINTEGER('D', "delay", &initial_delay,
1589                      "ms to wait before starting measurement after program start"),
1590         OPT_END()
1591         };
1592         const char * const stat_usage[] = {
1593                 "perf stat [<options>] [<command>]",
1594                 NULL
1595         };
1596         int status = -ENOMEM, run_idx;
1597         const char *mode;
1598
1599         setlocale(LC_ALL, "");
1600
1601         evsel_list = perf_evlist__new();
1602         if (evsel_list == NULL)
1603                 return -ENOMEM;
1604
1605         argc = parse_options(argc, argv, options, stat_usage,
1606                 PARSE_OPT_STOP_AT_NON_OPTION);
1607
1608         output = stderr;
1609         if (output_name && strcmp(output_name, "-"))
1610                 output = NULL;
1611
1612         if (output_name && output_fd) {
1613                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1614                 usage_with_options(stat_usage, options);
1615         }
1616
1617         if (output_fd < 0) {
1618                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1619                 usage_with_options(stat_usage, options);
1620         }
1621
1622         if (!output) {
1623                 struct timespec tm;
1624                 mode = append_file ? "a" : "w";
1625
1626                 output = fopen(output_name, mode);
1627                 if (!output) {
1628                         perror("failed to create output file");
1629                         return -1;
1630                 }
1631                 clock_gettime(CLOCK_REALTIME, &tm);
1632                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1633         } else if (output_fd > 0) {
1634                 mode = append_file ? "a" : "w";
1635                 output = fdopen(output_fd, mode);
1636                 if (!output) {
1637                         perror("Failed opening logfd");
1638                         return -errno;
1639                 }
1640         }
1641
1642         if (csv_sep) {
1643                 csv_output = true;
1644                 if (!strcmp(csv_sep, "\\t"))
1645                         csv_sep = "\t";
1646         } else
1647                 csv_sep = DEFAULT_SEPARATOR;
1648
1649         /*
1650          * let the spreadsheet do the pretty-printing
1651          */
1652         if (csv_output) {
1653                 /* User explicitly passed -B? */
1654                 if (big_num_opt == 1) {
1655                         fprintf(stderr, "-B option not supported with -x\n");
1656                         usage_with_options(stat_usage, options);
1657                 } else /* Nope, so disable big number formatting */
1658                         big_num = false;
1659         } else if (big_num_opt == 0) /* User passed --no-big-num */
1660                 big_num = false;
1661
1662         if (!argc && perf_target__none(&target))
1663                 usage_with_options(stat_usage, options);
1664
1665         if (run_count < 0) {
1666                 usage_with_options(stat_usage, options);
1667         } else if (run_count == 0) {
1668                 forever = true;
1669                 run_count = 1;
1670         }
1671
1672         /* no_aggr, cgroup are for system-wide only */
1673         if ((aggr_mode != AGGR_GLOBAL || nr_cgroups)
1674              && !perf_target__has_cpu(&target)) {
1675                 fprintf(stderr, "both cgroup and no-aggregation "
1676                         "modes only available in system-wide mode\n");
1677
1678                 usage_with_options(stat_usage, options);
1679                 return -1;
1680         }
1681
1682         if (add_default_attributes())
1683                 goto out;
1684
1685         perf_target__validate(&target);
1686
1687         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1688                 if (perf_target__has_task(&target))
1689                         pr_err("Problems finding threads of monitor\n");
1690                 if (perf_target__has_cpu(&target))
1691                         perror("failed to parse CPUs map");
1692
1693                 usage_with_options(stat_usage, options);
1694                 return -1;
1695         }
1696         if (interval && interval < 100) {
1697                 pr_err("print interval must be >= 100ms\n");
1698                 usage_with_options(stat_usage, options);
1699                 return -1;
1700         }
1701
1702         if (perf_evlist__alloc_stats(evsel_list, interval))
1703                 goto out_free_maps;
1704
1705         if (perf_stat_init_aggr_mode())
1706                 goto out;
1707
1708         /*
1709          * We dont want to block the signals - that would cause
1710          * child tasks to inherit that and Ctrl-C would not work.
1711          * What we want is for Ctrl-C to work in the exec()-ed
1712          * task, but being ignored by perf stat itself:
1713          */
1714         atexit(sig_atexit);
1715         if (!forever)
1716                 signal(SIGINT,  skip_signal);
1717         signal(SIGCHLD, skip_signal);
1718         signal(SIGALRM, skip_signal);
1719         signal(SIGABRT, skip_signal);
1720
1721         status = 0;
1722         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1723                 if (run_count != 1 && verbose)
1724                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1725                                 run_idx + 1);
1726
1727                 status = run_perf_stat(argc, argv);
1728                 if (forever && status != -1) {
1729                         print_stat(argc, argv);
1730                         perf_stat__reset_stats(evsel_list);
1731                 }
1732         }
1733
1734         if (!forever && status != -1 && !interval)
1735                 print_stat(argc, argv);
1736
1737         perf_evlist__free_stats(evsel_list);
1738 out_free_maps:
1739         perf_evlist__delete_maps(evsel_list);
1740 out:
1741         perf_evlist__delete(evsel_list);
1742         return status;
1743 }