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         char name[25];
 710
 711         aggr_printout(evsel, cpu, nr);
 712
 713         scnprintf(name, sizeof(name), "%s%s",
 714                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
 715         fprintf(output, fmt, msecs, csv_sep, name);
 716
 717         if (evsel->cgrp)
 718                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 719
 720         if (csv_output || interval)
 721                 return;
 722
 723         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 724                 fprintf(output, " # %8.3f CPUs utilized          ",
 725                         avg / avg_stats(&walltime_nsecs_stats));
 726         else
 727                 fprintf(output, "                                   ");
 728 }
 729
 730 /* used for get_ratio_color() */
 731 enum grc_type {
 732         GRC_STALLED_CYCLES_FE,
 733         GRC_STALLED_CYCLES_BE,
 734         GRC_CACHE_MISSES,
 735         GRC_MAX_NR
 736 };
 737
 738 static const char *get_ratio_color(enum grc_type type, double ratio)
 739 {
 740         static const double grc_table[GRC_MAX_NR][3] = {
 741                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
 742                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
 743                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
 744         };
 745         const char *color = PERF_COLOR_NORMAL;
 746
 747         if (ratio > grc_table[type][0])
 748                 color = PERF_COLOR_RED;
 749         else if (ratio > grc_table[type][1])
 750                 color = PERF_COLOR_MAGENTA;
 751         else if (ratio > grc_table[type][2])
 752                 color = PERF_COLOR_YELLOW;
 753
 754         return color;
 755 }
 756
 757 static void print_stalled_cycles_frontend(int cpu,
 758                                           struct perf_evsel *evsel
 759                                           __maybe_unused, double avg)
 760 {
 761         double total, ratio = 0.0;
 762         const char *color;
 763
 764         total = avg_stats(&runtime_cycles_stats[cpu]);
 765
 766         if (total)
 767                 ratio = avg / total * 100.0;
 768
 769         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
 770
 771         fprintf(output, " #  ");
 772         color_fprintf(output, color, "%6.2f%%", ratio);
 773         fprintf(output, " frontend cycles idle   ");
 774 }
 775
 776 static void print_stalled_cycles_backend(int cpu,
 777                                          struct perf_evsel *evsel
 778                                          __maybe_unused, double avg)
 779 {
 780         double total, ratio = 0.0;
 781         const char *color;
 782
 783         total = avg_stats(&runtime_cycles_stats[cpu]);
 784
 785         if (total)
 786                 ratio = avg / total * 100.0;
 787
 788         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
 789
 790         fprintf(output, " #  ");
 791         color_fprintf(output, color, "%6.2f%%", ratio);
 792         fprintf(output, " backend  cycles idle   ");
 793 }
 794
 795 static void print_branch_misses(int cpu,
 796                                 struct perf_evsel *evsel __maybe_unused,
 797                                 double avg)
 798 {
 799         double total, ratio = 0.0;
 800         const char *color;
 801
 802         total = avg_stats(&runtime_branches_stats[cpu]);
 803
 804         if (total)
 805                 ratio = avg / total * 100.0;
 806
 807         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 808
 809         fprintf(output, " #  ");
 810         color_fprintf(output, color, "%6.2f%%", ratio);
 811         fprintf(output, " of all branches        ");
 812 }
 813
 814 static void print_l1_dcache_misses(int cpu,
 815                                    struct perf_evsel *evsel __maybe_unused,
 816                                    double avg)
 817 {
 818         double total, ratio = 0.0;
 819         const char *color;
 820
 821         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
 822
 823         if (total)
 824                 ratio = avg / total * 100.0;
 825
 826         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 827
 828         fprintf(output, " #  ");
 829         color_fprintf(output, color, "%6.2f%%", ratio);
 830         fprintf(output, " of all L1-dcache hits  ");
 831 }
 832
 833 static void print_l1_icache_misses(int cpu,
 834                                    struct perf_evsel *evsel __maybe_unused,
 835                                    double avg)
 836 {
 837         double total, ratio = 0.0;
 838         const char *color;
 839
 840         total = avg_stats(&runtime_l1_icache_stats[cpu]);
 841
 842         if (total)
 843                 ratio = avg / total * 100.0;
 844
 845         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 846
 847         fprintf(output, " #  ");
 848         color_fprintf(output, color, "%6.2f%%", ratio);
 849         fprintf(output, " of all L1-icache hits  ");
 850 }
 851
 852 static void print_dtlb_cache_misses(int cpu,
 853                                     struct perf_evsel *evsel __maybe_unused,
 854                                     double avg)
 855 {
 856         double total, ratio = 0.0;
 857         const char *color;
 858
 859         total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
 860
 861         if (total)
 862                 ratio = avg / total * 100.0;
 863
 864         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 865
 866         fprintf(output, " #  ");
 867         color_fprintf(output, color, "%6.2f%%", ratio);
 868         fprintf(output, " of all dTLB cache hits ");
 869 }
 870
 871 static void print_itlb_cache_misses(int cpu,
 872                                     struct perf_evsel *evsel __maybe_unused,
 873                                     double avg)
 874 {
 875         double total, ratio = 0.0;
 876         const char *color;
 877
 878         total = avg_stats(&runtime_itlb_cache_stats[cpu]);
 879
 880         if (total)
 881                 ratio = avg / total * 100.0;
 882
 883         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 884
 885         fprintf(output, " #  ");
 886         color_fprintf(output, color, "%6.2f%%", ratio);
 887         fprintf(output, " of all iTLB cache hits ");
 888 }
 889
 890 static void print_ll_cache_misses(int cpu,
 891                                   struct perf_evsel *evsel __maybe_unused,
 892                                   double avg)
 893 {
 894         double total, ratio = 0.0;
 895         const char *color;
 896
 897         total = avg_stats(&runtime_ll_cache_stats[cpu]);
 898
 899         if (total)
 900                 ratio = avg / total * 100.0;
 901
 902         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 903
 904         fprintf(output, " #  ");
 905         color_fprintf(output, color, "%6.2f%%", ratio);
 906         fprintf(output, " of all LL-cache hits   ");
 907 }
 908
 909 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 910 {
 911         double total, ratio = 0.0, total2;
 912         const char *fmt;
 913
 914         if (csv_output)
 915                 fmt = "%.0f%s%s";
 916         else if (big_num)
 917                 fmt = "%'18.0f%s%-25s";
 918         else
 919                 fmt = "%18.0f%s%-25s";
 920
 921         aggr_printout(evsel, cpu, nr);
 922
 923         if (aggr_mode == AGGR_GLOBAL)
 924                 cpu = 0;
 925
 926         fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
 927
 928         if (evsel->cgrp)
 929                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 930
 931         if (csv_output || interval)
 932                 return;
 933
 934         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 935                 total = avg_stats(&runtime_cycles_stats[cpu]);
 936                 if (total) {
 937                         ratio = avg / total;
 938                         fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
 939                 }
 940                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
 941                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
 942
 943                 if (total && avg) {
 944                         ratio = total / avg;
 945                         fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
 946                 }
 947
 948         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 949                         runtime_branches_stats[cpu].n != 0) {
 950                 print_branch_misses(cpu, evsel, avg);
 951         } else if (
 952                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 953                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
 954                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 955                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 956                         runtime_l1_dcache_stats[cpu].n != 0) {
 957                 print_l1_dcache_misses(cpu, evsel, avg);
 958         } else if (
 959                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 960                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
 961                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 962                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 963                         runtime_l1_icache_stats[cpu].n != 0) {
 964                 print_l1_icache_misses(cpu, evsel, avg);
 965         } else if (
 966                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 967                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
 968                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 969                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 970                         runtime_dtlb_cache_stats[cpu].n != 0) {
 971                 print_dtlb_cache_misses(cpu, evsel, avg);
 972         } else if (
 973                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 974                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
 975                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 976                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 977                         runtime_itlb_cache_stats[cpu].n != 0) {
 978                 print_itlb_cache_misses(cpu, evsel, avg);
 979         } else if (
 980                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 981                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
 982                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 983                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 984                         runtime_ll_cache_stats[cpu].n != 0) {
 985                 print_ll_cache_misses(cpu, evsel, avg);
 986         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
 987                         runtime_cacherefs_stats[cpu].n != 0) {
 988                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
 989
 990                 if (total)
 991                         ratio = avg * 100 / total;
 992
 993                 fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
 994
 995         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
 996                 print_stalled_cycles_frontend(cpu, evsel, avg);
 997         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
 998                 print_stalled_cycles_backend(cpu, evsel, avg);
 999         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1000                 total = avg_stats(&runtime_nsecs_stats[cpu]);
1001
1002                 if (total) {
1003                         ratio = avg / total;
1004                         fprintf(output, " # %8.3f GHz                    ", ratio);
1005                 }
1006         } else if (transaction_run &&
1007                    perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1008                 total = avg_stats(&runtime_cycles_stats[cpu]);
1009                 if (total)
1010                         fprintf(output,
1011                                 " #   %5.2f%% transactional cycles   ",
1012                                 100.0 * (avg / total));
1013         } else if (transaction_run &&
1014                    perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1015                 total = avg_stats(&runtime_cycles_stats[cpu]);
1016                 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1017                 if (total2 < avg)
1018                         total2 = avg;
1019                 if (total)
1020                         fprintf(output,
1021                                 " #   %5.2f%% aborted cycles         ",
1022                                 100.0 * ((total2-avg) / total));
1023         } else if (transaction_run &&
1024                    perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1025                    avg > 0 &&
1026                    runtime_cycles_in_tx_stats[cpu].n != 0) {
1027                 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1028
1029                 if (total)
1030                         ratio = total / avg;
1031
1032                 fprintf(output, " # %8.0f cycles / transaction   ", ratio);
1033         } else if (transaction_run &&
1034                    perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1035                    avg > 0 &&
1036                    runtime_cycles_in_tx_stats[cpu].n != 0) {
1037                 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1038
1039                 if (total)
1040                         ratio = total / avg;
1041
1042                 fprintf(output, " # %8.0f cycles / elision       ", ratio);
1043         } else if (runtime_nsecs_stats[cpu].n != 0) {
1044                 char unit = 'M';
1045
1046                 total = avg_stats(&runtime_nsecs_stats[cpu]);
1047
1048                 if (total)
1049                         ratio = 1000.0 * avg / total;
1050                 if (ratio < 0.001) {
1051                         ratio *= 1000;
1052                         unit = 'K';
1053                 }
1054
1055                 fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
1056         } else {
1057                 fprintf(output, "                                   ");
1058         }
1059 }
1060
1061 static void print_aggr(char *prefix)
1062 {
1063         struct perf_evsel *counter;
1064         int cpu, cpu2, s, s2, id, nr;
1065         u64 ena, run, val;
1066
1067         if (!(aggr_map || aggr_get_id))
1068                 return;
1069
1070         for (s = 0; s < aggr_map->nr; s++) {
1071                 id = aggr_map->map[s];
1072                 list_for_each_entry(counter, &evsel_list->entries, node) {
1073                         val = ena = run = 0;
1074                         nr = 0;
1075                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1076                                 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1077                                 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1078                                 if (s2 != id)
1079                                         continue;
1080                                 val += counter->counts->cpu[cpu].val;
1081                                 ena += counter->counts->cpu[cpu].ena;
1082                                 run += counter->counts->cpu[cpu].run;
1083                                 nr++;
1084                         }
1085                         if (prefix)
1086                                 fprintf(output, "%s", prefix);
1087
1088                         if (run == 0 || ena == 0) {
1089                                 aggr_printout(counter, id, nr);
1090
1091                                 fprintf(output, "%*s%s%*s",
1092                                         csv_output ? 0 : 18,
1093                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1094                                         csv_sep,
1095                                         csv_output ? 0 : -24,
1096                                         perf_evsel__name(counter));
1097
1098                                 if (counter->cgrp)
1099                                         fprintf(output, "%s%s",
1100                                                 csv_sep, counter->cgrp->name);
1101
1102                                 fputc('\n', output);
1103                                 continue;
1104                         }
1105
1106                         if (nsec_counter(counter))
1107                                 nsec_printout(id, nr, counter, val);
1108                         else
1109                                 abs_printout(id, nr, counter, val);
1110
1111                         if (!csv_output) {
1112                                 print_noise(counter, 1.0);
1113
1114                                 if (run != ena)
1115                                         fprintf(output, "  (%.2f%%)",
1116                                                 100.0 * run / ena);
1117                         }
1118                         fputc('\n', output);
1119                 }
1120         }
1121 }
1122
1123 /*
1124  * Print out the results of a single counter:
1125  * aggregated counts in system-wide mode
1126  */
1127 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1128 {
1129         struct perf_stat *ps = counter->priv;
1130         double avg = avg_stats(&ps->res_stats[0]);
1131         int scaled = counter->counts->scaled;
1132
1133         if (prefix)
1134                 fprintf(output, "%s", prefix);
1135
1136         if (scaled == -1) {
1137                 fprintf(output, "%*s%s%*s",
1138                         csv_output ? 0 : 18,
1139                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1140                         csv_sep,
1141                         csv_output ? 0 : -24,
1142                         perf_evsel__name(counter));
1143
1144                 if (counter->cgrp)
1145                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1146
1147                 fputc('\n', output);
1148                 return;
1149         }
1150
1151         if (nsec_counter(counter))
1152                 nsec_printout(-1, 0, counter, avg);
1153         else
1154                 abs_printout(-1, 0, counter, avg);
1155
1156         print_noise(counter, avg);
1157
1158         if (csv_output) {
1159                 fputc('\n', output);
1160                 return;
1161         }
1162
1163         if (scaled) {
1164                 double avg_enabled, avg_running;
1165
1166                 avg_enabled = avg_stats(&ps->res_stats[1]);
1167                 avg_running = avg_stats(&ps->res_stats[2]);
1168
1169                 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1170         }
1171         fprintf(output, "\n");
1172 }
1173
1174 /*
1175  * Print out the results of a single counter:
1176  * does not use aggregated count in system-wide
1177  */
1178 static void print_counter(struct perf_evsel *counter, char *prefix)
1179 {
1180         u64 ena, run, val;
1181         int cpu;
1182
1183         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1184                 val = counter->counts->cpu[cpu].val;
1185                 ena = counter->counts->cpu[cpu].ena;
1186                 run = counter->counts->cpu[cpu].run;
1187
1188                 if (prefix)
1189                         fprintf(output, "%s", prefix);
1190
1191                 if (run == 0 || ena == 0) {
1192                         fprintf(output, "CPU%*d%s%*s%s%*s",
1193                                 csv_output ? 0 : -4,
1194                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1195                                 csv_output ? 0 : 18,
1196                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1197                                 csv_sep,
1198                                 csv_output ? 0 : -24,
1199                                 perf_evsel__name(counter));
1200
1201                         if (counter->cgrp)
1202                                 fprintf(output, "%s%s",
1203                                         csv_sep, counter->cgrp->name);
1204
1205                         fputc('\n', output);
1206                         continue;
1207                 }
1208
1209                 if (nsec_counter(counter))
1210                         nsec_printout(cpu, 0, counter, val);
1211                 else
1212                         abs_printout(cpu, 0, counter, val);
1213
1214                 if (!csv_output) {
1215                         print_noise(counter, 1.0);
1216
1217                         if (run != ena)
1218                                 fprintf(output, "  (%.2f%%)",
1219                                         100.0 * run / ena);
1220                 }
1221                 fputc('\n', output);
1222         }
1223 }
1224
1225 static void print_stat(int argc, const char **argv)
1226 {
1227         struct perf_evsel *counter;
1228         int i;
1229
1230         fflush(stdout);
1231
1232         if (!csv_output) {
1233                 fprintf(output, "\n");
1234                 fprintf(output, " Performance counter stats for ");
1235                 if (target.system_wide)
1236                         fprintf(output, "\'system wide");
1237                 else if (target.cpu_list)
1238                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1239                 else if (!perf_target__has_task(&target)) {
1240                         fprintf(output, "\'%s", argv[0]);
1241                         for (i = 1; i < argc; i++)
1242                                 fprintf(output, " %s", argv[i]);
1243                 } else if (target.pid)
1244                         fprintf(output, "process id \'%s", target.pid);
1245                 else
1246                         fprintf(output, "thread id \'%s", target.tid);
1247
1248                 fprintf(output, "\'");
1249                 if (run_count > 1)
1250                         fprintf(output, " (%d runs)", run_count);
1251                 fprintf(output, ":\n\n");
1252         }
1253
1254         switch (aggr_mode) {
1255         case AGGR_CORE:
1256         case AGGR_SOCKET:
1257                 print_aggr(NULL);
1258                 break;
1259         case AGGR_GLOBAL:
1260                 list_for_each_entry(counter, &evsel_list->entries, node)
1261                         print_counter_aggr(counter, NULL);
1262                 break;
1263         case AGGR_NONE:
1264                 list_for_each_entry(counter, &evsel_list->entries, node)
1265                         print_counter(counter, NULL);
1266                 break;
1267         default:
1268                 break;
1269         }
1270
1271         if (!csv_output) {
1272                 if (!null_run)
1273                         fprintf(output, "\n");
1274                 fprintf(output, " %17.9f seconds time elapsed",
1275                                 avg_stats(&walltime_nsecs_stats)/1e9);
1276                 if (run_count > 1) {
1277                         fprintf(output, "                                        ");
1278                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1279                                         avg_stats(&walltime_nsecs_stats));
1280                 }
1281                 fprintf(output, "\n\n");
1282         }
1283 }
1284
1285 static volatile int signr = -1;
1286
1287 static void skip_signal(int signo)
1288 {
1289         if ((child_pid == -1) || interval)
1290                 done = 1;
1291
1292         signr = signo;
1293         /*
1294          * render child_pid harmless
1295          * won't send SIGTERM to a random
1296          * process in case of race condition
1297          * and fast PID recycling
1298          */
1299         child_pid = -1;
1300 }
1301
1302 static void sig_atexit(void)
1303 {
1304         sigset_t set, oset;
1305
1306         /*
1307          * avoid race condition with SIGCHLD handler
1308          * in skip_signal() which is modifying child_pid
1309          * goal is to avoid send SIGTERM to a random
1310          * process
1311          */
1312         sigemptyset(&set);
1313         sigaddset(&set, SIGCHLD);
1314         sigprocmask(SIG_BLOCK, &set, &oset);
1315
1316         if (child_pid != -1)
1317                 kill(child_pid, SIGTERM);
1318
1319         sigprocmask(SIG_SETMASK, &oset, NULL);
1320
1321         if (signr == -1)
1322                 return;
1323
1324         signal(signr, SIG_DFL);
1325         kill(getpid(), signr);
1326 }
1327
1328 static int stat__set_big_num(const struct option *opt __maybe_unused,
1329                              const char *s __maybe_unused, int unset)
1330 {
1331         big_num_opt = unset ? 0 : 1;
1332         return 0;
1333 }
1334
1335 static int perf_stat_init_aggr_mode(void)
1336 {
1337         switch (aggr_mode) {
1338         case AGGR_SOCKET:
1339                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1340                         perror("cannot build socket map");
1341                         return -1;
1342                 }
1343                 aggr_get_id = cpu_map__get_socket;
1344                 break;
1345         case AGGR_CORE:
1346                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1347                         perror("cannot build core map");
1348                         return -1;
1349                 }
1350                 aggr_get_id = cpu_map__get_core;
1351                 break;
1352         case AGGR_NONE:
1353         case AGGR_GLOBAL:
1354         default:
1355                 break;
1356         }
1357         return 0;
1358 }
1359
1360 static int setup_events(const char * const *attrs, unsigned len)
1361 {
1362         unsigned i;
1363
1364         for (i = 0; i < len; i++) {
1365                 if (parse_events(evsel_list, attrs[i]))
1366                         return -1;
1367         }
1368         return 0;
1369 }
1370
1371 /*
1372  * Add default attributes, if there were no attributes specified or
1373  * if -d/--detailed, -d -d or -d -d -d is used:
1374  */
1375 static int add_default_attributes(void)
1376 {
1377         struct perf_event_attr default_attrs[] = {
1378
1379   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1380   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1381   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1382   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1383
1384   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1385   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1386   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1387   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1388   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1389   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1390
1391 };
1392
1393 /*
1394  * Detailed stats (-d), covering the L1 and last level data caches:
1395  */
1396         struct perf_event_attr detailed_attrs[] = {
1397
1398   { .type = PERF_TYPE_HW_CACHE,
1399     .config =
1400          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1401         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1402         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1403
1404   { .type = PERF_TYPE_HW_CACHE,
1405     .config =
1406          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1407         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1408         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1409
1410   { .type = PERF_TYPE_HW_CACHE,
1411     .config =
1412          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1413         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1414         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1415
1416   { .type = PERF_TYPE_HW_CACHE,
1417     .config =
1418          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1419         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1420         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1421 };
1422
1423 /*
1424  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1425  */
1426         struct perf_event_attr very_detailed_attrs[] = {
1427
1428   { .type = PERF_TYPE_HW_CACHE,
1429     .config =
1430          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1431         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1432         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1433
1434   { .type = PERF_TYPE_HW_CACHE,
1435     .config =
1436          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1437         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1438         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1439
1440   { .type = PERF_TYPE_HW_CACHE,
1441     .config =
1442          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1443         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1444         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1445
1446   { .type = PERF_TYPE_HW_CACHE,
1447     .config =
1448          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1449         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1450         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1451
1452   { .type = PERF_TYPE_HW_CACHE,
1453     .config =
1454          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1455         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1456         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1457
1458   { .type = PERF_TYPE_HW_CACHE,
1459     .config =
1460          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1461         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1462         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1463
1464 };
1465
1466 /*
1467  * Very, very detailed stats (-d -d -d), adding prefetch events:
1468  */
1469         struct perf_event_attr very_very_detailed_attrs[] = {
1470
1471   { .type = PERF_TYPE_HW_CACHE,
1472     .config =
1473          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1474         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1475         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1476
1477   { .type = PERF_TYPE_HW_CACHE,
1478     .config =
1479          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1480         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1481         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1482 };
1483
1484         /* Set attrs if no event is selected and !null_run: */
1485         if (null_run)
1486                 return 0;
1487
1488         if (transaction_run) {
1489                 int err;
1490                 if (pmu_have_event("cpu", "cycles-ct") &&
1491                     pmu_have_event("cpu", "el-start"))
1492                         err = setup_events(transaction_attrs,
1493                                         ARRAY_SIZE(transaction_attrs));
1494                 else
1495                         err = setup_events(transaction_limited_attrs,
1496                                  ARRAY_SIZE(transaction_limited_attrs));
1497                 if (err < 0) {
1498                         fprintf(stderr, "Cannot set up transaction events\n");
1499                         return -1;
1500                 }
1501                 return 0;
1502         }
1503
1504         if (!evsel_list->nr_entries) {
1505                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1506                         return -1;
1507         }
1508
1509         /* Detailed events get appended to the event list: */
1510
1511         if (detailed_run <  1)
1512                 return 0;
1513
1514         /* Append detailed run extra attributes: */
1515         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1516                 return -1;
1517
1518         if (detailed_run < 2)
1519                 return 0;
1520
1521         /* Append very detailed run extra attributes: */
1522         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1523                 return -1;
1524
1525         if (detailed_run < 3)
1526                 return 0;
1527
1528         /* Append very, very detailed run extra attributes: */
1529         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1530 }
1531
1532 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1533 {
1534         bool append_file = false;
1535         int output_fd = 0;
1536         const char *output_name = NULL;
1537         const struct option options[] = {
1538         OPT_BOOLEAN('T', "transaction", &transaction_run,
1539                     "hardware transaction statistics"),
1540         OPT_CALLBACK('e', "event", &evsel_list, "event",
1541                      "event selector. use 'perf list' to list available events",
1542                      parse_events_option),
1543         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1544                      "event filter", parse_filter),
1545         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1546                     "child tasks do not inherit counters"),
1547         OPT_STRING('p', "pid", &target.pid, "pid",
1548                    "stat events on existing process id"),
1549         OPT_STRING('t', "tid", &target.tid, "tid",
1550                    "stat events on existing thread id"),
1551         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1552                     "system-wide collection from all CPUs"),
1553         OPT_BOOLEAN('g', "group", &group,
1554                     "put the counters into a counter group"),
1555         OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1556         OPT_INCR('v', "verbose", &verbose,
1557                     "be more verbose (show counter open errors, etc)"),
1558         OPT_INTEGER('r', "repeat", &run_count,
1559                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1560         OPT_BOOLEAN('n', "null", &null_run,
1561                     "null run - dont start any counters"),
1562         OPT_INCR('d', "detailed", &detailed_run,
1563                     "detailed run - start a lot of events"),
1564         OPT_BOOLEAN('S', "sync", &sync_run,
1565                     "call sync() before starting a run"),
1566         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1567                            "print large numbers with thousands\' separators",
1568                            stat__set_big_num),
1569         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1570                     "list of cpus to monitor in system-wide"),
1571         OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1572                     "disable CPU count aggregation", AGGR_NONE),
1573         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1574                    "print counts with custom separator"),
1575         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1576                      "monitor event in cgroup name only", parse_cgroups),
1577         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1578         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1579         OPT_INTEGER(0, "log-fd", &output_fd,
1580                     "log output to fd, instead of stderr"),
1581         OPT_STRING(0, "pre", &pre_cmd, "command",
1582                         "command to run prior to the measured command"),
1583         OPT_STRING(0, "post", &post_cmd, "command",
1584                         "command to run after to the measured command"),
1585         OPT_UINTEGER('I', "interval-print", &interval,
1586                     "print counts at regular interval in ms (>= 100)"),
1587         OPT_SET_UINT(0, "per-socket", &aggr_mode,
1588                      "aggregate counts per processor socket", AGGR_SOCKET),
1589         OPT_SET_UINT(0, "per-core", &aggr_mode,
1590                      "aggregate counts per physical processor core", AGGR_CORE),
1591         OPT_UINTEGER('D', "delay", &initial_delay,
1592                      "ms to wait before starting measurement after program start"),
1593         OPT_END()
1594         };
1595         const char * const stat_usage[] = {
1596                 "perf stat [<options>] [<command>]",
1597                 NULL
1598         };
1599         int status = -EINVAL, run_idx;
1600         const char *mode;
1601
1602         setlocale(LC_ALL, "");
1603
1604         evsel_list = perf_evlist__new();
1605         if (evsel_list == NULL)
1606                 return -ENOMEM;
1607
1608         argc = parse_options(argc, argv, options, stat_usage,
1609                 PARSE_OPT_STOP_AT_NON_OPTION);
1610
1611         output = stderr;
1612         if (output_name && strcmp(output_name, "-"))
1613                 output = NULL;
1614
1615         if (output_name && output_fd) {
1616                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1617                 parse_options_usage(stat_usage, options, "o", 1);
1618                 parse_options_usage(NULL, options, "log-fd", 0);
1619                 goto out;
1620         }
1621
1622         if (output_fd < 0) {
1623                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1624                 parse_options_usage(stat_usage, options, "log-fd", 0);
1625                 goto out;
1626         }
1627
1628         if (!output) {
1629                 struct timespec tm;
1630                 mode = append_file ? "a" : "w";
1631
1632                 output = fopen(output_name, mode);
1633                 if (!output) {
1634                         perror("failed to create output file");
1635                         return -1;
1636                 }
1637                 clock_gettime(CLOCK_REALTIME, &tm);
1638                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1639         } else if (output_fd > 0) {
1640                 mode = append_file ? "a" : "w";
1641                 output = fdopen(output_fd, mode);
1642                 if (!output) {
1643                         perror("Failed opening logfd");
1644                         return -errno;
1645                 }
1646         }
1647
1648         if (csv_sep) {
1649                 csv_output = true;
1650                 if (!strcmp(csv_sep, "\\t"))
1651                         csv_sep = "\t";
1652         } else
1653                 csv_sep = DEFAULT_SEPARATOR;
1654
1655         /*
1656          * let the spreadsheet do the pretty-printing
1657          */
1658         if (csv_output) {
1659                 /* User explicitly passed -B? */
1660                 if (big_num_opt == 1) {
1661                         fprintf(stderr, "-B option not supported with -x\n");
1662                         parse_options_usage(stat_usage, options, "B", 1);
1663                         parse_options_usage(NULL, options, "x", 1);
1664                         goto out;
1665                 } else /* Nope, so disable big number formatting */
1666                         big_num = false;
1667         } else if (big_num_opt == 0) /* User passed --no-big-num */
1668                 big_num = false;
1669
1670         if (!argc && perf_target__none(&target))
1671                 usage_with_options(stat_usage, options);
1672
1673         if (run_count < 0) {
1674                 pr_err("Run count must be a positive number\n");
1675                 parse_options_usage(stat_usage, options, "r", 1);
1676                 goto out;
1677         } else if (run_count == 0) {
1678                 forever = true;
1679                 run_count = 1;
1680         }
1681
1682         /* no_aggr, cgroup are for system-wide only */
1683         if ((aggr_mode != AGGR_GLOBAL || nr_cgroups)
1684              && !perf_target__has_cpu(&target)) {
1685                 fprintf(stderr, "both cgroup and no-aggregation "
1686                         "modes only available in system-wide mode\n");
1687
1688                 parse_options_usage(stat_usage, options, "G", 1);
1689                 parse_options_usage(NULL, options, "A", 1);
1690                 parse_options_usage(NULL, options, "a", 1);
1691                 goto out;
1692         }
1693
1694         if (add_default_attributes())
1695                 goto out;
1696
1697         perf_target__validate(&target);
1698
1699         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1700                 if (perf_target__has_task(&target)) {
1701                         pr_err("Problems finding threads of monitor\n");
1702                         parse_options_usage(stat_usage, options, "p", 1);
1703                         parse_options_usage(NULL, options, "t", 1);
1704                 } else if (perf_target__has_cpu(&target)) {
1705                         perror("failed to parse CPUs map");
1706                         parse_options_usage(stat_usage, options, "C", 1);
1707                         parse_options_usage(NULL, options, "a", 1);
1708                 }
1709                 goto out;
1710         }
1711         if (interval && interval < 100) {
1712                 pr_err("print interval must be >= 100ms\n");
1713                 parse_options_usage(stat_usage, options, "I", 1);
1714                 goto out_free_maps;
1715         }
1716
1717         if (perf_evlist__alloc_stats(evsel_list, interval))
1718                 goto out_free_maps;
1719
1720         if (perf_stat_init_aggr_mode())
1721                 goto out_free_maps;
1722
1723         /*
1724          * We dont want to block the signals - that would cause
1725          * child tasks to inherit that and Ctrl-C would not work.
1726          * What we want is for Ctrl-C to work in the exec()-ed
1727          * task, but being ignored by perf stat itself:
1728          */
1729         atexit(sig_atexit);
1730         if (!forever)
1731                 signal(SIGINT,  skip_signal);
1732         signal(SIGCHLD, skip_signal);
1733         signal(SIGALRM, skip_signal);
1734         signal(SIGABRT, skip_signal);
1735
1736         status = 0;
1737         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1738                 if (run_count != 1 && verbose)
1739                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1740                                 run_idx + 1);
1741
1742                 status = run_perf_stat(argc, argv);
1743                 if (forever && status != -1) {
1744                         print_stat(argc, argv);
1745                         perf_stat__reset_stats(evsel_list);
1746                 }
1747         }
1748
1749         if (!forever && status != -1 && !interval)
1750                 print_stat(argc, argv);
1751
1752         perf_evlist__free_stats(evsel_list);
1753 out_free_maps:
1754         perf_evlist__delete_maps(evsel_list);
1755 out:
1756         perf_evlist__delete(evsel_list);
1757         return status;
1758 }