tools/perf/builtin-stat.c

   1 /*
   2  * builtin-stat.c
   3  *
   4  * Builtin stat command: Give a precise performance counters summary
   5  * overview about any workload, CPU or specific PID.
   6  *
   7  * Sample output:
   8
   9    $ perf stat ./hackbench 10
  10
  11   Time: 0.118
  12
  13   Performance counter stats for './hackbench 10':
  14
  15        1708.761321 task-clock                #   11.037 CPUs utilized
  16             41,190 context-switches          #    0.024 M/sec
  17              6,735 CPU-migrations            #    0.004 M/sec
  18             17,318 page-faults               #    0.010 M/sec
  19      5,205,202,243 cycles                    #    3.046 GHz
  20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  22      2,603,501,247 instructions              #    0.50  insns per cycle
  23                                              #    1.48  stalled cycles per insn
  24        484,357,498 branches                  #  283.455 M/sec
  25          6,388,934 branch-misses             #    1.32% of all branches
  26
  27         0.154822978  seconds time elapsed
  28
  29  *
  30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  31  *
  32  * Improvements and fixes by:
  33  *
  34  *   Arjan van de Ven <arjan@linux.intel.com>
  35  *   Yanmin Zhang <yanmin.zhang@intel.com>
  36  *   Wu Fengguang <fengguang.wu@intel.com>
  37  *   Mike Galbraith <efault@gmx.de>
  38  *   Paul Mackerras <paulus@samba.org>
  39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  40  *
  41  * Released under the GPL v2. (and only v2, not any later version)
  42  */
  43
  44 #include "perf.h"
  45 #include "builtin.h"
  46 #include "util/cgroup.h"
  47 #include "util/util.h"
  48 #include "util/parse-options.h"
  49 #include "util/parse-events.h"
  50 #include "util/pmu.h"
  51 #include "util/event.h"
  52 #include "util/evlist.h"
  53 #include "util/evsel.h"
  54 #include "util/debug.h"
  55 #include "util/color.h"
  56 #include "util/stat.h"
  57 #include "util/header.h"
  58 #include "util/cpumap.h"
  59 #include "util/thread.h"
  60 #include "util/thread_map.h"
  61
  62 #include <stdlib.h>
  63 #include <sys/prctl.h>
  64 #include <locale.h>
  65
  66 #define DEFAULT_SEPARATOR       " "
  67 #define CNTR_NOT_SUPPORTED      "<not supported>"
  68 #define CNTR_NOT_COUNTED        "<not counted>"
  69
  70 static void print_stat(int argc, const char **argv);
  71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
  72 static void print_counter(struct perf_evsel *counter, char *prefix);
  73 static void print_aggr(char *prefix);
  74
  75 /* Default events used for perf stat -T */
  76 static const char * const transaction_attrs[] = {
  77         "task-clock",
  78         "{"
  79         "instructions,"
  80         "cycles,"
  81         "cpu/cycles-t/,"
  82         "cpu/tx-start/,"
  83         "cpu/el-start/,"
  84         "cpu/cycles-ct/"
  85         "}"
  86 };
  87
  88 /* More limited version when the CPU does not have all events. */
  89 static const char * const transaction_limited_attrs[] = {
  90         "task-clock",
  91         "{"
  92         "instructions,"
  93         "cycles,"
  94         "cpu/cycles-t/,"
  95         "cpu/tx-start/"
  96         "}"
  97 };
  98
  99 /* must match transaction_attrs and the beginning limited_attrs */
 100 enum {
 101         T_TASK_CLOCK,
 102         T_INSTRUCTIONS,
 103         T_CYCLES,
 104         T_CYCLES_IN_TX,
 105         T_TRANSACTION_START,
 106         T_ELISION_START,
 107         T_CYCLES_IN_TX_CP,
 108 };
 109
 110 static struct perf_evlist       *evsel_list;
 111
 112 static struct target target = {
 113         .uid    = UINT_MAX,
 114 };
 115
 116 enum aggr_mode {
 117         AGGR_NONE,
 118         AGGR_GLOBAL,
 119         AGGR_SOCKET,
 120         AGGR_CORE,
 121 };
 122
 123 static int                      run_count                       =  1;
 124 static bool                     no_inherit                      = false;
 125 static bool                     scale                           =  true;
 126 static enum aggr_mode           aggr_mode                       = AGGR_GLOBAL;
 127 static volatile pid_t           child_pid                       = -1;
 128 static bool                     null_run                        =  false;
 129 static int                      detailed_run                    =  0;
 130 static bool                     transaction_run;
 131 static bool                     big_num                         =  true;
 132 static int                      big_num_opt                     =  -1;
 133 static const char               *csv_sep                        = NULL;
 134 static bool                     csv_output                      = false;
 135 static bool                     group                           = false;
 136 static FILE                     *output                         = NULL;
 137 static const char               *pre_cmd                        = NULL;
 138 static const char               *post_cmd                       = NULL;
 139 static bool                     sync_run                        = false;
 140 static unsigned int             interval                        = 0;
 141 static unsigned int             initial_delay                   = 0;
 142 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 143 static bool                     forever                         = false;
 144 static struct timespec          ref_time;
 145 static struct cpu_map           *aggr_map;
 146 static int                      (*aggr_get_id)(struct cpu_map *m, int cpu);
 147
 148 static volatile int done = 0;
 149
 150 struct perf_stat {
 151         struct stats      res_stats[3];
 152 };
 153
 154 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 155                                  struct timespec *b)
 156 {
 157         r->tv_sec = a->tv_sec - b->tv_sec;
 158         if (a->tv_nsec < b->tv_nsec) {
 159                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 160                 r->tv_sec--;
 161         } else {
 162                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 163         }
 164 }
 165
 166 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
 167 {
 168         return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
 169 }
 170
 171 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
 172 {
 173         return perf_evsel__cpus(evsel)->nr;
 174 }
 175
 176 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
 177 {
 178         int i;
 179         struct perf_stat *ps = evsel->priv;
 180
 181         for (i = 0; i < 3; i++)
 182                 init_stats(&ps->res_stats[i]);
 183 }
 184
 185 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 186 {
 187         evsel->priv = zalloc(sizeof(struct perf_stat));
 188         if (evsel->priv == NULL)
 189                 return -ENOMEM;
 190         perf_evsel__reset_stat_priv(evsel);
 191         return 0;
 192 }
 193
 194 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 195 {
 196         zfree(&evsel->priv);
 197 }
 198
 199 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
 200 {
 201         void *addr;
 202         size_t sz;
 203
 204         sz = sizeof(*evsel->counts) +
 205              (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
 206
 207         addr = zalloc(sz);
 208         if (!addr)
 209                 return -ENOMEM;
 210
 211         evsel->prev_raw_counts =  addr;
 212
 213         return 0;
 214 }
 215
 216 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
 217 {
 218         zfree(&evsel->prev_raw_counts);
 219 }
 220
 221 static void perf_evlist__free_stats(struct perf_evlist *evlist)
 222 {
 223         struct perf_evsel *evsel;
 224
 225         evlist__for_each(evlist, evsel) {
 226                 perf_evsel__free_stat_priv(evsel);
 227                 perf_evsel__free_counts(evsel);
 228                 perf_evsel__free_prev_raw_counts(evsel);
 229         }
 230 }
 231
 232 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
 233 {
 234         struct perf_evsel *evsel;
 235
 236         evlist__for_each(evlist, evsel) {
 237                 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
 238                     perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
 239                     (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
 240                         goto out_free;
 241         }
 242
 243         return 0;
 244
 245 out_free:
 246         perf_evlist__free_stats(evlist);
 247         return -1;
 248 }
 249
 250 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
 251 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
 252 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
 253 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
 254 static struct stats runtime_branches_stats[MAX_NR_CPUS];
 255 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
 256 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
 257 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
 258 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
 259 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
 260 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
 261 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
 262 static struct stats walltime_nsecs_stats;
 263 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
 264 static struct stats runtime_elision_stats[MAX_NR_CPUS];
 265
 266 static void perf_stat__reset_stats(struct perf_evlist *evlist)
 267 {
 268         struct perf_evsel *evsel;
 269
 270         evlist__for_each(evlist, evsel) {
 271                 perf_evsel__reset_stat_priv(evsel);
 272                 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
 273         }
 274
 275         memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
 276         memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
 277         memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
 278         memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
 279         memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
 280         memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
 281         memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
 282         memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
 283         memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
 284         memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
 285         memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
 286         memset(runtime_cycles_in_tx_stats, 0,
 287                         sizeof(runtime_cycles_in_tx_stats));
 288         memset(runtime_transaction_stats, 0,
 289                 sizeof(runtime_transaction_stats));
 290         memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
 291         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
 292 }
 293
 294 static int create_perf_stat_counter(struct perf_evsel *evsel)
 295 {
 296         struct perf_event_attr *attr = &evsel->attr;
 297
 298         if (scale)
 299                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 300                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 301
 302         attr->inherit = !no_inherit;
 303
 304         if (target__has_cpu(&target))
 305                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 306
 307         if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
 308                 attr->disabled = 1;
 309                 if (!initial_delay)
 310                         attr->enable_on_exec = 1;
 311         }
 312
 313         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 314 }
 315
 316 /*
 317  * Does the counter have nsecs as a unit?
 318  */
 319 static inline int nsec_counter(struct perf_evsel *evsel)
 320 {
 321         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 322             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 323                 return 1;
 324
 325         return 0;
 326 }
 327
 328 static struct perf_evsel *nth_evsel(int n)
 329 {
 330         static struct perf_evsel **array;
 331         static int array_len;
 332         struct perf_evsel *ev;
 333         int j;
 334
 335         /* Assumes this only called when evsel_list does not change anymore. */
 336         if (!array) {
 337                 evlist__for_each(evsel_list, ev)
 338                         array_len++;
 339                 array = malloc(array_len * sizeof(void *));
 340                 if (!array)
 341                         exit(ENOMEM);
 342                 j = 0;
 343                 evlist__for_each(evsel_list, ev)
 344                         array[j++] = ev;
 345         }
 346         if (n < array_len)
 347                 return array[n];
 348         return NULL;
 349 }
 350
 351 /*
 352  * Update various tracking values we maintain to print
 353  * more semantic information such as miss/hit ratios,
 354  * instruction rates, etc:
 355  */
 356 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
 357 {
 358         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 359                 update_stats(&runtime_nsecs_stats[0], count[0]);
 360         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 361                 update_stats(&runtime_cycles_stats[0], count[0]);
 362         else if (transaction_run &&
 363                  perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
 364                 update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
 365         else if (transaction_run &&
 366                  perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
 367                 update_stats(&runtime_transaction_stats[0], count[0]);
 368         else if (transaction_run &&
 369                  perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
 370                 update_stats(&runtime_elision_stats[0], count[0]);
 371         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
 372                 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
 373         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 374                 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
 375         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 376                 update_stats(&runtime_branches_stats[0], count[0]);
 377         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 378                 update_stats(&runtime_cacherefs_stats[0], count[0]);
 379         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
 380                 update_stats(&runtime_l1_dcache_stats[0], count[0]);
 381         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
 382                 update_stats(&runtime_l1_icache_stats[0], count[0]);
 383         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
 384                 update_stats(&runtime_ll_cache_stats[0], count[0]);
 385         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
 386                 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
 387         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
 388                 update_stats(&runtime_itlb_cache_stats[0], count[0]);
 389 }
 390
 391 /*
 392  * Read out the results of a single counter:
 393  * aggregate counts across CPUs in system-wide mode
 394  */
 395 static int read_counter_aggr(struct perf_evsel *counter)
 396 {
 397         struct perf_stat *ps = counter->priv;
 398         u64 *count = counter->counts->aggr.values;
 399         int i;
 400
 401         if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
 402                                thread_map__nr(evsel_list->threads), scale) < 0)
 403                 return -1;
 404
 405         for (i = 0; i < 3; i++)
 406                 update_stats(&ps->res_stats[i], count[i]);
 407
 408         if (verbose) {
 409                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 410                         perf_evsel__name(counter), count[0], count[1], count[2]);
 411         }
 412
 413         /*
 414          * Save the full runtime - to allow normalization during printout:
 415          */
 416         update_shadow_stats(counter, count);
 417
 418         return 0;
 419 }
 420
 421 /*
 422  * Read out the results of a single counter:
 423  * do not aggregate counts across CPUs in system-wide mode
 424  */
 425 static int read_counter(struct perf_evsel *counter)
 426 {
 427         u64 *count;
 428         int cpu;
 429
 430         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
 431                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 432                         return -1;
 433
 434                 count = counter->counts->cpu[cpu].values;
 435
 436                 update_shadow_stats(counter, count);
 437         }
 438
 439         return 0;
 440 }
 441
 442 static void print_interval(void)
 443 {
 444         static int num_print_interval;
 445         struct perf_evsel *counter;
 446         struct perf_stat *ps;
 447         struct timespec ts, rs;
 448         char prefix[64];
 449
 450         if (aggr_mode == AGGR_GLOBAL) {
 451                 evlist__for_each(evsel_list, counter) {
 452                         ps = counter->priv;
 453                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
 454                         read_counter_aggr(counter);
 455                 }
 456         } else  {
 457                 evlist__for_each(evsel_list, counter) {
 458                         ps = counter->priv;
 459                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
 460                         read_counter(counter);
 461                 }
 462         }
 463
 464         clock_gettime(CLOCK_MONOTONIC, &ts);
 465         diff_timespec(&rs, &ts, &ref_time);
 466         sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
 467
 468         if (num_print_interval == 0 && !csv_output) {
 469                 switch (aggr_mode) {
 470                 case AGGR_SOCKET:
 471                         fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
 472                         break;
 473                 case AGGR_CORE:
 474                         fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
 475                         break;
 476                 case AGGR_NONE:
 477                         fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
 478                         break;
 479                 case AGGR_GLOBAL:
 480                 default:
 481                         fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
 482                 }
 483         }
 484
 485         if (++num_print_interval == 25)
 486                 num_print_interval = 0;
 487
 488         switch (aggr_mode) {
 489         case AGGR_CORE:
 490         case AGGR_SOCKET:
 491                 print_aggr(prefix);
 492                 break;
 493         case AGGR_NONE:
 494                 evlist__for_each(evsel_list, counter)
 495                         print_counter(counter, prefix);
 496                 break;
 497         case AGGR_GLOBAL:
 498         default:
 499                 evlist__for_each(evsel_list, counter)
 500                         print_counter_aggr(counter, prefix);
 501         }
 502
 503         fflush(output);
 504 }
 505
 506 static void handle_initial_delay(void)
 507 {
 508         struct perf_evsel *counter;
 509
 510         if (initial_delay) {
 511                 const int ncpus = cpu_map__nr(evsel_list->cpus),
 512                         nthreads = thread_map__nr(evsel_list->threads);
 513
 514                 usleep(initial_delay * 1000);
 515                 evlist__for_each(evsel_list, counter)
 516                         perf_evsel__enable(counter, ncpus, nthreads);
 517         }
 518 }
 519
 520 static volatile int workload_exec_errno;
 521
 522 /*
 523  * perf_evlist__prepare_workload will send a SIGUSR1
 524  * if the fork fails, since we asked by setting its
 525  * want_signal to true.
 526  */
 527 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 528                                         void *ucontext __maybe_unused)
 529 {
 530         workload_exec_errno = info->si_value.sival_int;
 531 }
 532
 533 static int __run_perf_stat(int argc, const char **argv)
 534 {
 535         char msg[512];
 536         unsigned long long t0, t1;
 537         struct perf_evsel *counter;
 538         struct timespec ts;
 539         size_t l;
 540         int status = 0;
 541         const bool forks = (argc > 0);
 542
 543         if (interval) {
 544                 ts.tv_sec  = interval / 1000;
 545                 ts.tv_nsec = (interval % 1000) * 1000000;
 546         } else {
 547                 ts.tv_sec  = 1;
 548                 ts.tv_nsec = 0;
 549         }
 550
 551         if (forks) {
 552                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
 553                                                   workload_exec_failed_signal) < 0) {
 554                         perror("failed to prepare workload");
 555                         return -1;
 556                 }
 557                 child_pid = evsel_list->workload.pid;
 558         }
 559
 560         if (group)
 561                 perf_evlist__set_leader(evsel_list);
 562
 563         evlist__for_each(evsel_list, counter) {
 564                 if (create_perf_stat_counter(counter) < 0) {
 565                         /*
 566                          * PPC returns ENXIO for HW counters until 2.6.37
 567                          * (behavior changed with commit b0a873e).
 568                          */
 569                         if (errno == EINVAL || errno == ENOSYS ||
 570                             errno == ENOENT || errno == EOPNOTSUPP ||
 571                             errno == ENXIO) {
 572                                 if (verbose)
 573                                         ui__warning("%s event is not supported by the kernel.\n",
 574                                                     perf_evsel__name(counter));
 575                                 counter->supported = false;
 576                                 continue;
 577                         }
 578
 579                         perf_evsel__open_strerror(counter, &target,
 580                                                   errno, msg, sizeof(msg));
 581                         ui__error("%s\n", msg);
 582
 583                         if (child_pid != -1)
 584                                 kill(child_pid, SIGTERM);
 585
 586                         return -1;
 587                 }
 588                 counter->supported = true;
 589
 590                 l = strlen(counter->unit);
 591                 if (l > unit_width)
 592                         unit_width = l;
 593         }
 594
 595         if (perf_evlist__apply_filters(evsel_list)) {
 596                 error("failed to set filter with %d (%s)\n", errno,
 597                         strerror_r(errno, msg, sizeof(msg)));
 598                 return -1;
 599         }
 600
 601         /*
 602          * Enable counters and exec the command:
 603          */
 604         t0 = rdclock();
 605         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 606
 607         if (forks) {
 608                 perf_evlist__start_workload(evsel_list);
 609                 handle_initial_delay();
 610
 611                 if (interval) {
 612                         while (!waitpid(child_pid, &status, WNOHANG)) {
 613                                 nanosleep(&ts, NULL);
 614                                 print_interval();
 615                         }
 616                 }
 617                 wait(&status);
 618
 619                 if (workload_exec_errno) {
 620                         const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 621                         pr_err("Workload failed: %s\n", emsg);
 622                         return -1;
 623                 }
 624
 625                 if (WIFSIGNALED(status))
 626                         psignal(WTERMSIG(status), argv[0]);
 627         } else {
 628                 handle_initial_delay();
 629                 while (!done) {
 630                         nanosleep(&ts, NULL);
 631                         if (interval)
 632                                 print_interval();
 633                 }
 634         }
 635
 636         t1 = rdclock();
 637
 638         update_stats(&walltime_nsecs_stats, t1 - t0);
 639
 640         if (aggr_mode == AGGR_GLOBAL) {
 641                 evlist__for_each(evsel_list, counter) {
 642                         read_counter_aggr(counter);
 643                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 644                                              thread_map__nr(evsel_list->threads));
 645                 }
 646         } else {
 647                 evlist__for_each(evsel_list, counter) {
 648                         read_counter(counter);
 649                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
 650                 }
 651         }
 652
 653         return WEXITSTATUS(status);
 654 }
 655
 656 static int run_perf_stat(int argc, const char **argv)
 657 {
 658         int ret;
 659
 660         if (pre_cmd) {
 661                 ret = system(pre_cmd);
 662                 if (ret)
 663                         return ret;
 664         }
 665
 666         if (sync_run)
 667                 sync();
 668
 669         ret = __run_perf_stat(argc, argv);
 670         if (ret)
 671                 return ret;
 672
 673         if (post_cmd) {
 674                 ret = system(post_cmd);
 675                 if (ret)
 676                         return ret;
 677         }
 678
 679         return ret;
 680 }
 681
 682 static void print_noise_pct(double total, double avg)
 683 {
 684         double pct = rel_stddev_stats(total, avg);
 685
 686         if (csv_output)
 687                 fprintf(output, "%s%.2f%%", csv_sep, pct);
 688         else if (pct)
 689                 fprintf(output, "  ( +-%6.2f%% )", pct);
 690 }
 691
 692 static void print_noise(struct perf_evsel *evsel, double avg)
 693 {
 694         struct perf_stat *ps;
 695
 696         if (run_count == 1)
 697                 return;
 698
 699         ps = evsel->priv;
 700         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 701 }
 702
 703 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 704 {
 705         switch (aggr_mode) {
 706         case AGGR_CORE:
 707                 fprintf(output, "S%d-C%*d%s%*d%s",
 708                         cpu_map__id_to_socket(id),
 709                         csv_output ? 0 : -8,
 710                         cpu_map__id_to_cpu(id),
 711                         csv_sep,
 712                         csv_output ? 0 : 4,
 713                         nr,
 714                         csv_sep);
 715                 break;
 716         case AGGR_SOCKET:
 717                 fprintf(output, "S%*d%s%*d%s",
 718                         csv_output ? 0 : -5,
 719                         id,
 720                         csv_sep,
 721                         csv_output ? 0 : 4,
 722                         nr,
 723                         csv_sep);
 724                         break;
 725         case AGGR_NONE:
 726                 fprintf(output, "CPU%*d%s",
 727                         csv_output ? 0 : -4,
 728                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 729                 break;
 730         case AGGR_GLOBAL:
 731         default:
 732                 break;
 733         }
 734 }
 735
 736 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 737 {
 738         double msecs = avg / 1e6;
 739         const char *fmt_v, *fmt_n;
 740         char name[25];
 741
 742         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
 743         fmt_n = csv_output ? "%s" : "%-25s";
 744
 745         aggr_printout(evsel, id, nr);
 746
 747         scnprintf(name, sizeof(name), "%s%s",
 748                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
 749
 750         fprintf(output, fmt_v, msecs, csv_sep);
 751
 752         if (csv_output)
 753                 fprintf(output, "%s%s", evsel->unit, csv_sep);
 754         else
 755                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
 756
 757         fprintf(output, fmt_n, name);
 758
 759         if (evsel->cgrp)
 760                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 761
 762         if (csv_output || interval)
 763                 return;
 764
 765         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 766                 fprintf(output, " # %8.3f CPUs utilized          ",
 767                         avg / avg_stats(&walltime_nsecs_stats));
 768         else
 769                 fprintf(output, "                                   ");
 770 }
 771
 772 /* used for get_ratio_color() */
 773 enum grc_type {
 774         GRC_STALLED_CYCLES_FE,
 775         GRC_STALLED_CYCLES_BE,
 776         GRC_CACHE_MISSES,
 777         GRC_MAX_NR
 778 };
 779
 780 static const char *get_ratio_color(enum grc_type type, double ratio)
 781 {
 782         static const double grc_table[GRC_MAX_NR][3] = {
 783                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
 784                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
 785                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
 786         };
 787         const char *color = PERF_COLOR_NORMAL;
 788
 789         if (ratio > grc_table[type][0])
 790                 color = PERF_COLOR_RED;
 791         else if (ratio > grc_table[type][1])
 792                 color = PERF_COLOR_MAGENTA;
 793         else if (ratio > grc_table[type][2])
 794                 color = PERF_COLOR_YELLOW;
 795
 796         return color;
 797 }
 798
 799 static void print_stalled_cycles_frontend(int cpu,
 800                                           struct perf_evsel *evsel
 801                                           __maybe_unused, double avg)
 802 {
 803         double total, ratio = 0.0;
 804         const char *color;
 805
 806         total = avg_stats(&runtime_cycles_stats[cpu]);
 807
 808         if (total)
 809                 ratio = avg / total * 100.0;
 810
 811         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
 812
 813         fprintf(output, " #  ");
 814         color_fprintf(output, color, "%6.2f%%", ratio);
 815         fprintf(output, " frontend cycles idle   ");
 816 }
 817
 818 static void print_stalled_cycles_backend(int cpu,
 819                                          struct perf_evsel *evsel
 820                                          __maybe_unused, double avg)
 821 {
 822         double total, ratio = 0.0;
 823         const char *color;
 824
 825         total = avg_stats(&runtime_cycles_stats[cpu]);
 826
 827         if (total)
 828                 ratio = avg / total * 100.0;
 829
 830         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
 831
 832         fprintf(output, " #  ");
 833         color_fprintf(output, color, "%6.2f%%", ratio);
 834         fprintf(output, " backend  cycles idle   ");
 835 }
 836
 837 static void print_branch_misses(int cpu,
 838                                 struct perf_evsel *evsel __maybe_unused,
 839                                 double avg)
 840 {
 841         double total, ratio = 0.0;
 842         const char *color;
 843
 844         total = avg_stats(&runtime_branches_stats[cpu]);
 845
 846         if (total)
 847                 ratio = avg / total * 100.0;
 848
 849         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 850
 851         fprintf(output, " #  ");
 852         color_fprintf(output, color, "%6.2f%%", ratio);
 853         fprintf(output, " of all branches        ");
 854 }
 855
 856 static void print_l1_dcache_misses(int cpu,
 857                                    struct perf_evsel *evsel __maybe_unused,
 858                                    double avg)
 859 {
 860         double total, ratio = 0.0;
 861         const char *color;
 862
 863         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
 864
 865         if (total)
 866                 ratio = avg / total * 100.0;
 867
 868         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 869
 870         fprintf(output, " #  ");
 871         color_fprintf(output, color, "%6.2f%%", ratio);
 872         fprintf(output, " of all L1-dcache hits  ");
 873 }
 874
 875 static void print_l1_icache_misses(int cpu,
 876                                    struct perf_evsel *evsel __maybe_unused,
 877                                    double avg)
 878 {
 879         double total, ratio = 0.0;
 880         const char *color;
 881
 882         total = avg_stats(&runtime_l1_icache_stats[cpu]);
 883
 884         if (total)
 885                 ratio = avg / total * 100.0;
 886
 887         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 888
 889         fprintf(output, " #  ");
 890         color_fprintf(output, color, "%6.2f%%", ratio);
 891         fprintf(output, " of all L1-icache hits  ");
 892 }
 893
 894 static void print_dtlb_cache_misses(int cpu,
 895                                     struct perf_evsel *evsel __maybe_unused,
 896                                     double avg)
 897 {
 898         double total, ratio = 0.0;
 899         const char *color;
 900
 901         total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
 902
 903         if (total)
 904                 ratio = avg / total * 100.0;
 905
 906         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 907
 908         fprintf(output, " #  ");
 909         color_fprintf(output, color, "%6.2f%%", ratio);
 910         fprintf(output, " of all dTLB cache hits ");
 911 }
 912
 913 static void print_itlb_cache_misses(int cpu,
 914                                     struct perf_evsel *evsel __maybe_unused,
 915                                     double avg)
 916 {
 917         double total, ratio = 0.0;
 918         const char *color;
 919
 920         total = avg_stats(&runtime_itlb_cache_stats[cpu]);
 921
 922         if (total)
 923                 ratio = avg / total * 100.0;
 924
 925         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 926
 927         fprintf(output, " #  ");
 928         color_fprintf(output, color, "%6.2f%%", ratio);
 929         fprintf(output, " of all iTLB cache hits ");
 930 }
 931
 932 static void print_ll_cache_misses(int cpu,
 933                                   struct perf_evsel *evsel __maybe_unused,
 934                                   double avg)
 935 {
 936         double total, ratio = 0.0;
 937         const char *color;
 938
 939         total = avg_stats(&runtime_ll_cache_stats[cpu]);
 940
 941         if (total)
 942                 ratio = avg / total * 100.0;
 943
 944         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
 945
 946         fprintf(output, " #  ");
 947         color_fprintf(output, color, "%6.2f%%", ratio);
 948         fprintf(output, " of all LL-cache hits   ");
 949 }
 950
 951 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 952 {
 953         double total, ratio = 0.0, total2;
 954         double sc =  evsel->scale;
 955         const char *fmt;
 956         int cpu = cpu_map__id_to_cpu(id);
 957
 958         if (csv_output) {
 959                 fmt = sc != 1.0 ?  "%.2f%s" : "%.0f%s";
 960         } else {
 961                 if (big_num)
 962                         fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
 963                 else
 964                         fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
 965         }
 966
 967         aggr_printout(evsel, id, nr);
 968
 969         if (aggr_mode == AGGR_GLOBAL)
 970                 cpu = 0;
 971
 972         fprintf(output, fmt, avg, csv_sep);
 973
 974         if (evsel->unit)
 975                 fprintf(output, "%-*s%s",
 976                         csv_output ? 0 : unit_width,
 977                         evsel->unit, csv_sep);
 978
 979         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
 980
 981         if (evsel->cgrp)
 982                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 983
 984         if (csv_output || interval)
 985                 return;
 986
 987         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 988                 total = avg_stats(&runtime_cycles_stats[cpu]);
 989                 if (total) {
 990                         ratio = avg / total;
 991                         fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
 992                 }
 993                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
 994                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
 995
 996                 if (total && avg) {
 997                         ratio = total / avg;
 998                         fprintf(output, "\n");
 999                         if (aggr_mode == AGGR_NONE)
1000                                 fprintf(output, "        ");
1001                         fprintf(output, "                                                  #   %5.2f  stalled cycles per insn", ratio);
1002                 }
1003
1004         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1005                         runtime_branches_stats[cpu].n != 0) {
1006                 print_branch_misses(cpu, evsel, avg);
1007         } else if (
1008                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1009                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
1010                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1011                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1012                         runtime_l1_dcache_stats[cpu].n != 0) {
1013                 print_l1_dcache_misses(cpu, evsel, avg);
1014         } else if (
1015                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1016                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
1017                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1018                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1019                         runtime_l1_icache_stats[cpu].n != 0) {
1020                 print_l1_icache_misses(cpu, evsel, avg);
1021         } else if (
1022                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1023                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
1024                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1025                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1026                         runtime_dtlb_cache_stats[cpu].n != 0) {
1027                 print_dtlb_cache_misses(cpu, evsel, avg);
1028         } else if (
1029                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1030                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
1031                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1032                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1033                         runtime_itlb_cache_stats[cpu].n != 0) {
1034                 print_itlb_cache_misses(cpu, evsel, avg);
1035         } else if (
1036                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1037                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
1038                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1039                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1040                         runtime_ll_cache_stats[cpu].n != 0) {
1041                 print_ll_cache_misses(cpu, evsel, avg);
1042         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1043                         runtime_cacherefs_stats[cpu].n != 0) {
1044                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1045
1046                 if (total)
1047                         ratio = avg * 100 / total;
1048
1049                 fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
1050
1051         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1052                 print_stalled_cycles_frontend(cpu, evsel, avg);
1053         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1054                 print_stalled_cycles_backend(cpu, evsel, avg);
1055         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1056                 total = avg_stats(&runtime_nsecs_stats[cpu]);
1057
1058                 if (total) {
1059                         ratio = avg / total;
1060                         fprintf(output, " # %8.3f GHz                    ", ratio);
1061                 }
1062         } else if (transaction_run &&
1063                    perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1064                 total = avg_stats(&runtime_cycles_stats[cpu]);
1065                 if (total)
1066                         fprintf(output,
1067                                 " #   %5.2f%% transactional cycles   ",
1068                                 100.0 * (avg / total));
1069         } else if (transaction_run &&
1070                    perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1071                 total = avg_stats(&runtime_cycles_stats[cpu]);
1072                 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1073                 if (total2 < avg)
1074                         total2 = avg;
1075                 if (total)
1076                         fprintf(output,
1077                                 " #   %5.2f%% aborted cycles         ",
1078                                 100.0 * ((total2-avg) / total));
1079         } else if (transaction_run &&
1080                    perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1081                    avg > 0 &&
1082                    runtime_cycles_in_tx_stats[cpu].n != 0) {
1083                 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1084
1085                 if (total)
1086                         ratio = total / avg;
1087
1088                 fprintf(output, " # %8.0f cycles / transaction   ", ratio);
1089         } else if (transaction_run &&
1090                    perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1091                    avg > 0 &&
1092                    runtime_cycles_in_tx_stats[cpu].n != 0) {
1093                 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1094
1095                 if (total)
1096                         ratio = total / avg;
1097
1098                 fprintf(output, " # %8.0f cycles / elision       ", ratio);
1099         } else if (runtime_nsecs_stats[cpu].n != 0) {
1100                 char unit = 'M';
1101
1102                 total = avg_stats(&runtime_nsecs_stats[cpu]);
1103
1104                 if (total)
1105                         ratio = 1000.0 * avg / total;
1106                 if (ratio < 0.001) {
1107                         ratio *= 1000;
1108                         unit = 'K';
1109                 }
1110
1111                 fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
1112         } else {
1113                 fprintf(output, "                                   ");
1114         }
1115 }
1116
1117 static void print_aggr(char *prefix)
1118 {
1119         struct perf_evsel *counter;
1120         int cpu, cpu2, s, s2, id, nr;
1121         double uval;
1122         u64 ena, run, val;
1123
1124         if (!(aggr_map || aggr_get_id))
1125                 return;
1126
1127         for (s = 0; s < aggr_map->nr; s++) {
1128                 id = aggr_map->map[s];
1129                 evlist__for_each(evsel_list, counter) {
1130                         val = ena = run = 0;
1131                         nr = 0;
1132                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1133                                 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1134                                 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1135                                 if (s2 != id)
1136                                         continue;
1137                                 val += counter->counts->cpu[cpu].val;
1138                                 ena += counter->counts->cpu[cpu].ena;
1139                                 run += counter->counts->cpu[cpu].run;
1140                                 nr++;
1141                         }
1142                         if (prefix)
1143                                 fprintf(output, "%s", prefix);
1144
1145                         if (run == 0 || ena == 0) {
1146                                 aggr_printout(counter, id, nr);
1147
1148                                 fprintf(output, "%*s%s",
1149                                         csv_output ? 0 : 18,
1150                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1151                                         csv_sep);
1152
1153                                 fprintf(output, "%-*s%s",
1154                                         csv_output ? 0 : unit_width,
1155                                         counter->unit, csv_sep);
1156
1157                                 fprintf(output, "%*s",
1158                                         csv_output ? 0 : -25,
1159                                         perf_evsel__name(counter));
1160
1161                                 if (counter->cgrp)
1162                                         fprintf(output, "%s%s",
1163                                                 csv_sep, counter->cgrp->name);
1164
1165                                 fputc('\n', output);
1166                                 continue;
1167                         }
1168                         uval = val * counter->scale;
1169
1170                         if (nsec_counter(counter))
1171                                 nsec_printout(id, nr, counter, uval);
1172                         else
1173                                 abs_printout(id, nr, counter, uval);
1174
1175                         if (!csv_output) {
1176                                 print_noise(counter, 1.0);
1177
1178                                 if (run != ena)
1179                                         fprintf(output, "  (%.2f%%)",
1180                                                 100.0 * run / ena);
1181                         }
1182                         fputc('\n', output);
1183                 }
1184         }
1185 }
1186
1187 /*
1188  * Print out the results of a single counter:
1189  * aggregated counts in system-wide mode
1190  */
1191 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1192 {
1193         struct perf_stat *ps = counter->priv;
1194         double avg = avg_stats(&ps->res_stats[0]);
1195         int scaled = counter->counts->scaled;
1196         double uval;
1197
1198         if (prefix)
1199                 fprintf(output, "%s", prefix);
1200
1201         if (scaled == -1) {
1202                 fprintf(output, "%*s%s",
1203                         csv_output ? 0 : 18,
1204                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1205                         csv_sep);
1206                 fprintf(output, "%-*s%s",
1207                         csv_output ? 0 : unit_width,
1208                         counter->unit, csv_sep);
1209                 fprintf(output, "%*s",
1210                         csv_output ? 0 : -25,
1211                         perf_evsel__name(counter));
1212
1213                 if (counter->cgrp)
1214                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1215
1216                 fputc('\n', output);
1217                 return;
1218         }
1219
1220         uval = avg * counter->scale;
1221
1222         if (nsec_counter(counter))
1223                 nsec_printout(-1, 0, counter, uval);
1224         else
1225                 abs_printout(-1, 0, counter, uval);
1226
1227         print_noise(counter, avg);
1228
1229         if (csv_output) {
1230                 fputc('\n', output);
1231                 return;
1232         }
1233
1234         if (scaled) {
1235                 double avg_enabled, avg_running;
1236
1237                 avg_enabled = avg_stats(&ps->res_stats[1]);
1238                 avg_running = avg_stats(&ps->res_stats[2]);
1239
1240                 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1241         }
1242         fprintf(output, "\n");
1243 }
1244
1245 /*
1246  * Print out the results of a single counter:
1247  * does not use aggregated count in system-wide
1248  */
1249 static void print_counter(struct perf_evsel *counter, char *prefix)
1250 {
1251         u64 ena, run, val;
1252         double uval;
1253         int cpu;
1254
1255         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1256                 val = counter->counts->cpu[cpu].val;
1257                 ena = counter->counts->cpu[cpu].ena;
1258                 run = counter->counts->cpu[cpu].run;
1259
1260                 if (prefix)
1261                         fprintf(output, "%s", prefix);
1262
1263                 if (run == 0 || ena == 0) {
1264                         fprintf(output, "CPU%*d%s%*s%s",
1265                                 csv_output ? 0 : -4,
1266                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1267                                 csv_output ? 0 : 18,
1268                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1269                                 csv_sep);
1270
1271                                 fprintf(output, "%-*s%s",
1272                                         csv_output ? 0 : unit_width,
1273                                         counter->unit, csv_sep);
1274
1275                                 fprintf(output, "%*s",
1276                                         csv_output ? 0 : -25,
1277                                         perf_evsel__name(counter));
1278
1279                         if (counter->cgrp)
1280                                 fprintf(output, "%s%s",
1281                                         csv_sep, counter->cgrp->name);
1282
1283                         fputc('\n', output);
1284                         continue;
1285                 }
1286
1287                 uval = val * counter->scale;
1288
1289                 if (nsec_counter(counter))
1290                         nsec_printout(cpu, 0, counter, uval);
1291                 else
1292                         abs_printout(cpu, 0, counter, uval);
1293
1294                 if (!csv_output) {
1295                         print_noise(counter, 1.0);
1296
1297                         if (run != ena)
1298                                 fprintf(output, "  (%.2f%%)",
1299                                         100.0 * run / ena);
1300                 }
1301                 fputc('\n', output);
1302         }
1303 }
1304
1305 static void print_stat(int argc, const char **argv)
1306 {
1307         struct perf_evsel *counter;
1308         int i;
1309
1310         fflush(stdout);
1311
1312         if (!csv_output) {
1313                 fprintf(output, "\n");
1314                 fprintf(output, " Performance counter stats for ");
1315                 if (target.system_wide)
1316                         fprintf(output, "\'system wide");
1317                 else if (target.cpu_list)
1318                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1319                 else if (!target__has_task(&target)) {
1320                         fprintf(output, "\'%s", argv[0]);
1321                         for (i = 1; i < argc; i++)
1322                                 fprintf(output, " %s", argv[i]);
1323                 } else if (target.pid)
1324                         fprintf(output, "process id \'%s", target.pid);
1325                 else
1326                         fprintf(output, "thread id \'%s", target.tid);
1327
1328                 fprintf(output, "\'");
1329                 if (run_count > 1)
1330                         fprintf(output, " (%d runs)", run_count);
1331                 fprintf(output, ":\n\n");
1332         }
1333
1334         switch (aggr_mode) {
1335         case AGGR_CORE:
1336         case AGGR_SOCKET:
1337                 print_aggr(NULL);
1338                 break;
1339         case AGGR_GLOBAL:
1340                 evlist__for_each(evsel_list, counter)
1341                         print_counter_aggr(counter, NULL);
1342                 break;
1343         case AGGR_NONE:
1344                 evlist__for_each(evsel_list, counter)
1345                         print_counter(counter, NULL);
1346                 break;
1347         default:
1348                 break;
1349         }
1350
1351         if (!csv_output) {
1352                 if (!null_run)
1353                         fprintf(output, "\n");
1354                 fprintf(output, " %17.9f seconds time elapsed",
1355                                 avg_stats(&walltime_nsecs_stats)/1e9);
1356                 if (run_count > 1) {
1357                         fprintf(output, "                                        ");
1358                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1359                                         avg_stats(&walltime_nsecs_stats));
1360                 }
1361                 fprintf(output, "\n\n");
1362         }
1363 }
1364
1365 static volatile int signr = -1;
1366
1367 static void skip_signal(int signo)
1368 {
1369         if ((child_pid == -1) || interval)
1370                 done = 1;
1371
1372         signr = signo;
1373         /*
1374          * render child_pid harmless
1375          * won't send SIGTERM to a random
1376          * process in case of race condition
1377          * and fast PID recycling
1378          */
1379         child_pid = -1;
1380 }
1381
1382 static void sig_atexit(void)
1383 {
1384         sigset_t set, oset;
1385
1386         /*
1387          * avoid race condition with SIGCHLD handler
1388          * in skip_signal() which is modifying child_pid
1389          * goal is to avoid send SIGTERM to a random
1390          * process
1391          */
1392         sigemptyset(&set);
1393         sigaddset(&set, SIGCHLD);
1394         sigprocmask(SIG_BLOCK, &set, &oset);
1395
1396         if (child_pid != -1)
1397                 kill(child_pid, SIGTERM);
1398
1399         sigprocmask(SIG_SETMASK, &oset, NULL);
1400
1401         if (signr == -1)
1402                 return;
1403
1404         signal(signr, SIG_DFL);
1405         kill(getpid(), signr);
1406 }
1407
1408 static int stat__set_big_num(const struct option *opt __maybe_unused,
1409                              const char *s __maybe_unused, int unset)
1410 {
1411         big_num_opt = unset ? 0 : 1;
1412         return 0;
1413 }
1414
1415 static int perf_stat_init_aggr_mode(void)
1416 {
1417         switch (aggr_mode) {
1418         case AGGR_SOCKET:
1419                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1420                         perror("cannot build socket map");
1421                         return -1;
1422                 }
1423                 aggr_get_id = cpu_map__get_socket;
1424                 break;
1425         case AGGR_CORE:
1426                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1427                         perror("cannot build core map");
1428                         return -1;
1429                 }
1430                 aggr_get_id = cpu_map__get_core;
1431                 break;
1432         case AGGR_NONE:
1433         case AGGR_GLOBAL:
1434         default:
1435                 break;
1436         }
1437         return 0;
1438 }
1439
1440 static int setup_events(const char * const *attrs, unsigned len)
1441 {
1442         unsigned i;
1443
1444         for (i = 0; i < len; i++) {
1445                 if (parse_events(evsel_list, attrs[i]))
1446                         return -1;
1447         }
1448         return 0;
1449 }
1450
1451 /*
1452  * Add default attributes, if there were no attributes specified or
1453  * if -d/--detailed, -d -d or -d -d -d is used:
1454  */
1455 static int add_default_attributes(void)
1456 {
1457         struct perf_event_attr default_attrs[] = {
1458
1459   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1460   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1461   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1462   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1463
1464   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1465   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1466   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1467   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1468   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1469   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1470
1471 };
1472
1473 /*
1474  * Detailed stats (-d), covering the L1 and last level data caches:
1475  */
1476         struct perf_event_attr detailed_attrs[] = {
1477
1478   { .type = PERF_TYPE_HW_CACHE,
1479     .config =
1480          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1481         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1482         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1483
1484   { .type = PERF_TYPE_HW_CACHE,
1485     .config =
1486          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1487         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1488         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1489
1490   { .type = PERF_TYPE_HW_CACHE,
1491     .config =
1492          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1493         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1494         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1495
1496   { .type = PERF_TYPE_HW_CACHE,
1497     .config =
1498          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1499         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1500         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1501 };
1502
1503 /*
1504  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1505  */
1506         struct perf_event_attr very_detailed_attrs[] = {
1507
1508   { .type = PERF_TYPE_HW_CACHE,
1509     .config =
1510          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1511         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1512         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1513
1514   { .type = PERF_TYPE_HW_CACHE,
1515     .config =
1516          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1517         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1518         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1519
1520   { .type = PERF_TYPE_HW_CACHE,
1521     .config =
1522          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1523         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1524         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1525
1526   { .type = PERF_TYPE_HW_CACHE,
1527     .config =
1528          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1529         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1530         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1531
1532   { .type = PERF_TYPE_HW_CACHE,
1533     .config =
1534          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1535         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1536         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1537
1538   { .type = PERF_TYPE_HW_CACHE,
1539     .config =
1540          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1541         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1542         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1543
1544 };
1545
1546 /*
1547  * Very, very detailed stats (-d -d -d), adding prefetch events:
1548  */
1549         struct perf_event_attr very_very_detailed_attrs[] = {
1550
1551   { .type = PERF_TYPE_HW_CACHE,
1552     .config =
1553          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1554         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1555         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1556
1557   { .type = PERF_TYPE_HW_CACHE,
1558     .config =
1559          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1560         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1561         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1562 };
1563
1564         /* Set attrs if no event is selected and !null_run: */
1565         if (null_run)
1566                 return 0;
1567
1568         if (transaction_run) {
1569                 int err;
1570                 if (pmu_have_event("cpu", "cycles-ct") &&
1571                     pmu_have_event("cpu", "el-start"))
1572                         err = setup_events(transaction_attrs,
1573                                         ARRAY_SIZE(transaction_attrs));
1574                 else
1575                         err = setup_events(transaction_limited_attrs,
1576                                  ARRAY_SIZE(transaction_limited_attrs));
1577                 if (err < 0) {
1578                         fprintf(stderr, "Cannot set up transaction events\n");
1579                         return -1;
1580                 }
1581                 return 0;
1582         }
1583
1584         if (!evsel_list->nr_entries) {
1585                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1586                         return -1;
1587         }
1588
1589         /* Detailed events get appended to the event list: */
1590
1591         if (detailed_run <  1)
1592                 return 0;
1593
1594         /* Append detailed run extra attributes: */
1595         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1596                 return -1;
1597
1598         if (detailed_run < 2)
1599                 return 0;
1600
1601         /* Append very detailed run extra attributes: */
1602         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1603                 return -1;
1604
1605         if (detailed_run < 3)
1606                 return 0;
1607
1608         /* Append very, very detailed run extra attributes: */
1609         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1610 }
1611
1612 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1613 {
1614         bool append_file = false;
1615         int output_fd = 0;
1616         const char *output_name = NULL;
1617         const struct option options[] = {
1618         OPT_BOOLEAN('T', "transaction", &transaction_run,
1619                     "hardware transaction statistics"),
1620         OPT_CALLBACK('e', "event", &evsel_list, "event",
1621                      "event selector. use 'perf list' to list available events",
1622                      parse_events_option),
1623         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1624                      "event filter", parse_filter),
1625         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1626                     "child tasks do not inherit counters"),
1627         OPT_STRING('p', "pid", &target.pid, "pid",
1628                    "stat events on existing process id"),
1629         OPT_STRING('t', "tid", &target.tid, "tid",
1630                    "stat events on existing thread id"),
1631         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1632                     "system-wide collection from all CPUs"),
1633         OPT_BOOLEAN('g', "group", &group,
1634                     "put the counters into a counter group"),
1635         OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1636         OPT_INCR('v', "verbose", &verbose,
1637                     "be more verbose (show counter open errors, etc)"),
1638         OPT_INTEGER('r', "repeat", &run_count,
1639                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1640         OPT_BOOLEAN('n', "null", &null_run,
1641                     "null run - dont start any counters"),
1642         OPT_INCR('d', "detailed", &detailed_run,
1643                     "detailed run - start a lot of events"),
1644         OPT_BOOLEAN('S', "sync", &sync_run,
1645                     "call sync() before starting a run"),
1646         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1647                            "print large numbers with thousands\' separators",
1648                            stat__set_big_num),
1649         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1650                     "list of cpus to monitor in system-wide"),
1651         OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1652                     "disable CPU count aggregation", AGGR_NONE),
1653         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1654                    "print counts with custom separator"),
1655         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1656                      "monitor event in cgroup name only", parse_cgroups),
1657         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1658         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1659         OPT_INTEGER(0, "log-fd", &output_fd,
1660                     "log output to fd, instead of stderr"),
1661         OPT_STRING(0, "pre", &pre_cmd, "command",
1662                         "command to run prior to the measured command"),
1663         OPT_STRING(0, "post", &post_cmd, "command",
1664                         "command to run after to the measured command"),
1665         OPT_UINTEGER('I', "interval-print", &interval,
1666                     "print counts at regular interval in ms (>= 100)"),
1667         OPT_SET_UINT(0, "per-socket", &aggr_mode,
1668                      "aggregate counts per processor socket", AGGR_SOCKET),
1669         OPT_SET_UINT(0, "per-core", &aggr_mode,
1670                      "aggregate counts per physical processor core", AGGR_CORE),
1671         OPT_UINTEGER('D', "delay", &initial_delay,
1672                      "ms to wait before starting measurement after program start"),
1673         OPT_END()
1674         };
1675         const char * const stat_usage[] = {
1676                 "perf stat [<options>] [<command>]",
1677                 NULL
1678         };
1679         int status = -EINVAL, run_idx;
1680         const char *mode;
1681
1682         setlocale(LC_ALL, "");
1683
1684         evsel_list = perf_evlist__new();
1685         if (evsel_list == NULL)
1686                 return -ENOMEM;
1687
1688         argc = parse_options(argc, argv, options, stat_usage,
1689                 PARSE_OPT_STOP_AT_NON_OPTION);
1690
1691         output = stderr;
1692         if (output_name && strcmp(output_name, "-"))
1693                 output = NULL;
1694
1695         if (output_name && output_fd) {
1696                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1697                 parse_options_usage(stat_usage, options, "o", 1);
1698                 parse_options_usage(NULL, options, "log-fd", 0);
1699                 goto out;
1700         }
1701
1702         if (output_fd < 0) {
1703                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1704                 parse_options_usage(stat_usage, options, "log-fd", 0);
1705                 goto out;
1706         }
1707
1708         if (!output) {
1709                 struct timespec tm;
1710                 mode = append_file ? "a" : "w";
1711
1712                 output = fopen(output_name, mode);
1713                 if (!output) {
1714                         perror("failed to create output file");
1715                         return -1;
1716                 }
1717                 clock_gettime(CLOCK_REALTIME, &tm);
1718                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1719         } else if (output_fd > 0) {
1720                 mode = append_file ? "a" : "w";
1721                 output = fdopen(output_fd, mode);
1722                 if (!output) {
1723                         perror("Failed opening logfd");
1724                         return -errno;
1725                 }
1726         }
1727
1728         if (csv_sep) {
1729                 csv_output = true;
1730                 if (!strcmp(csv_sep, "\\t"))
1731                         csv_sep = "\t";
1732         } else
1733                 csv_sep = DEFAULT_SEPARATOR;
1734
1735         /*
1736          * let the spreadsheet do the pretty-printing
1737          */
1738         if (csv_output) {
1739                 /* User explicitly passed -B? */
1740                 if (big_num_opt == 1) {
1741                         fprintf(stderr, "-B option not supported with -x\n");
1742                         parse_options_usage(stat_usage, options, "B", 1);
1743                         parse_options_usage(NULL, options, "x", 1);
1744                         goto out;
1745                 } else /* Nope, so disable big number formatting */
1746                         big_num = false;
1747         } else if (big_num_opt == 0) /* User passed --no-big-num */
1748                 big_num = false;
1749
1750         if (!argc && target__none(&target))
1751                 usage_with_options(stat_usage, options);
1752
1753         if (run_count < 0) {
1754                 pr_err("Run count must be a positive number\n");
1755                 parse_options_usage(stat_usage, options, "r", 1);
1756                 goto out;
1757         } else if (run_count == 0) {
1758                 forever = true;
1759                 run_count = 1;
1760         }
1761
1762         /* no_aggr, cgroup are for system-wide only */
1763         if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1764             !target__has_cpu(&target)) {
1765                 fprintf(stderr, "both cgroup and no-aggregation "
1766                         "modes only available in system-wide mode\n");
1767
1768                 parse_options_usage(stat_usage, options, "G", 1);
1769                 parse_options_usage(NULL, options, "A", 1);
1770                 parse_options_usage(NULL, options, "a", 1);
1771                 goto out;
1772         }
1773
1774         if (add_default_attributes())
1775                 goto out;
1776
1777         target__validate(&target);
1778
1779         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1780                 if (target__has_task(&target)) {
1781                         pr_err("Problems finding threads of monitor\n");
1782                         parse_options_usage(stat_usage, options, "p", 1);
1783                         parse_options_usage(NULL, options, "t", 1);
1784                 } else if (target__has_cpu(&target)) {
1785                         perror("failed to parse CPUs map");
1786                         parse_options_usage(stat_usage, options, "C", 1);
1787                         parse_options_usage(NULL, options, "a", 1);
1788                 }
1789                 goto out;
1790         }
1791         if (interval && interval < 100) {
1792                 pr_err("print interval must be >= 100ms\n");
1793                 parse_options_usage(stat_usage, options, "I", 1);
1794                 goto out;
1795         }
1796
1797         if (perf_evlist__alloc_stats(evsel_list, interval))
1798                 goto out;
1799
1800         if (perf_stat_init_aggr_mode())
1801                 goto out;
1802
1803         /*
1804          * We dont want to block the signals - that would cause
1805          * child tasks to inherit that and Ctrl-C would not work.
1806          * What we want is for Ctrl-C to work in the exec()-ed
1807          * task, but being ignored by perf stat itself:
1808          */
1809         atexit(sig_atexit);
1810         if (!forever)
1811                 signal(SIGINT,  skip_signal);
1812         signal(SIGCHLD, skip_signal);
1813         signal(SIGALRM, skip_signal);
1814         signal(SIGABRT, skip_signal);
1815
1816         status = 0;
1817         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1818                 if (run_count != 1 && verbose)
1819                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1820                                 run_idx + 1);
1821
1822                 status = run_perf_stat(argc, argv);
1823                 if (forever && status != -1) {
1824                         print_stat(argc, argv);
1825                         perf_stat__reset_stats(evsel_list);
1826                 }
1827         }
1828
1829         if (!forever && status != -1 && !interval)
1830                 print_stat(argc, argv);
1831
1832         perf_evlist__free_stats(evsel_list);
1833 out:
1834         perf_evlist__delete(evsel_list);
1835         return status;
1836 }