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    Time: 0.104
  11
  12     Performance counter stats for '/home/mingo/hackbench':
  13
  14        1255.538611  task clock ticks     #      10.143 CPU utilization factor
  15              54011  context switches     #       0.043 M/sec
  16                385  CPU migrations       #       0.000 M/sec
  17              17755  pagefaults           #       0.014 M/sec
  18         3808323185  CPU cycles           #    3033.219 M/sec
  19         1575111190  instructions         #    1254.530 M/sec
  20           17367895  cache references     #      13.833 M/sec
  21            7674421  cache misses         #       6.112 M/sec
  22
  23     Wall-clock time elapsed:   123.786620 msecs
  24
  25  *
  26  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  27  *
  28  * Improvements and fixes by:
  29  *
  30  *   Arjan van de Ven <arjan@linux.intel.com>
  31  *   Yanmin Zhang <yanmin.zhang@intel.com>
  32  *   Wu Fengguang <fengguang.wu@intel.com>
  33  *   Mike Galbraith <efault@gmx.de>
  34  *   Paul Mackerras <paulus@samba.org>
  35  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  36  *
  37  * Released under the GPL v2. (and only v2, not any later version)
  38  */
  39
  40 #include "perf.h"
  41 #include "builtin.h"
  42 #include "util/util.h"
  43 #include "util/parse-options.h"
  44 #include "util/parse-events.h"
  45 #include "util/event.h"
  46 #include "util/evlist.h"
  47 #include "util/evsel.h"
  48 #include "util/debug.h"
  49 #include "util/color.h"
  50 #include "util/header.h"
  51 #include "util/cpumap.h"
  52 #include "util/thread.h"
  53 #include "util/thread_map.h"
  54
  55 #include <sys/prctl.h>
  56 #include <math.h>
  57 #include <locale.h>
  58
  59 #define DEFAULT_SEPARATOR       " "
  60
  61 static struct perf_event_attr default_attrs[] = {
  62
  63   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
  64   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
  65   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
  66   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
  67
  68   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
  69   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
  70   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
  71   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
  72   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
  73   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
  74
  75 };
  76
  77 /*
  78  * Detailed stats:
  79  */
  80 static struct perf_event_attr detailed_attrs[] = {
  81
  82   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
  83   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
  84   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
  85   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
  86
  87   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
  88   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
  89   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
  90   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
  91   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
  92   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
  93
  94   { .type = PERF_TYPE_HW_CACHE,
  95     .config =
  96          PERF_COUNT_HW_CACHE_L1D                <<  0  |
  97         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
  98         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
  99
 100   { .type = PERF_TYPE_HW_CACHE,
 101     .config =
 102          PERF_COUNT_HW_CACHE_L1D                <<  0  |
 103         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 104         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
 105
 106   { .type = PERF_TYPE_HW_CACHE,
 107     .config =
 108          PERF_COUNT_HW_CACHE_LL                 <<  0  |
 109         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 110         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
 111
 112   { .type = PERF_TYPE_HW_CACHE,
 113     .config =
 114          PERF_COUNT_HW_CACHE_LL                 <<  0  |
 115         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
 116         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
 117 };
 118
 119 struct perf_evlist              *evsel_list;
 120
 121 static bool                     system_wide                     =  false;
 122 static int                      run_idx                         =  0;
 123
 124 static int                      run_count                       =  1;
 125 static bool                     no_inherit                      = false;
 126 static bool                     scale                           =  true;
 127 static bool                     no_aggr                         = false;
 128 static pid_t                    target_pid                      = -1;
 129 static pid_t                    target_tid                      = -1;
 130 static pid_t                    child_pid                       = -1;
 131 static bool                     null_run                        =  false;
 132 static bool                     detailed_run                    =  false;
 133 static bool                     sync_run                        =  false;
 134 static bool                     big_num                         =  true;
 135 static int                      big_num_opt                     =  -1;
 136 static const char               *cpu_list;
 137 static const char               *csv_sep                        = NULL;
 138 static bool                     csv_output                      = false;
 139
 140 static volatile int done = 0;
 141
 142 struct stats
 143 {
 144         double n, mean, M2;
 145 };
 146
 147 struct perf_stat {
 148         struct stats      res_stats[3];
 149 };
 150
 151 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 152 {
 153         evsel->priv = zalloc(sizeof(struct perf_stat));
 154         return evsel->priv == NULL ? -ENOMEM : 0;
 155 }
 156
 157 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 158 {
 159         free(evsel->priv);
 160         evsel->priv = NULL;
 161 }
 162
 163 static void update_stats(struct stats *stats, u64 val)
 164 {
 165         double delta;
 166
 167         stats->n++;
 168         delta = val - stats->mean;
 169         stats->mean += delta / stats->n;
 170         stats->M2 += delta*(val - stats->mean);
 171 }
 172
 173 static double avg_stats(struct stats *stats)
 174 {
 175         return stats->mean;
 176 }
 177
 178 /*
 179  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 180  *
 181  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 182  * s^2 = -------------------------------
 183  *                  n - 1
 184  *
 185  * http://en.wikipedia.org/wiki/Stddev
 186  *
 187  * The std dev of the mean is related to the std dev by:
 188  *
 189  *             s
 190  * s_mean = -------
 191  *          sqrt(n)
 192  *
 193  */
 194 static double stddev_stats(struct stats *stats)
 195 {
 196         double variance = stats->M2 / (stats->n - 1);
 197         double variance_mean = variance / stats->n;
 198
 199         return sqrt(variance_mean);
 200 }
 201
 202 struct stats                    runtime_nsecs_stats[MAX_NR_CPUS];
 203 struct stats                    runtime_cycles_stats[MAX_NR_CPUS];
 204 struct stats                    runtime_stalled_cycles_stats[MAX_NR_CPUS];
 205 struct stats                    runtime_branches_stats[MAX_NR_CPUS];
 206 struct stats                    runtime_cacherefs_stats[MAX_NR_CPUS];
 207 struct stats                    runtime_l1_dcache_stats[MAX_NR_CPUS];
 208 struct stats                    walltime_nsecs_stats;
 209
 210 static int create_perf_stat_counter(struct perf_evsel *evsel)
 211 {
 212         struct perf_event_attr *attr = &evsel->attr;
 213
 214         if (scale)
 215                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 216                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 217
 218         attr->inherit = !no_inherit;
 219
 220         if (system_wide)
 221                 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false);
 222
 223         if (target_pid == -1 && target_tid == -1) {
 224                 attr->disabled = 1;
 225                 attr->enable_on_exec = 1;
 226         }
 227
 228         return perf_evsel__open_per_thread(evsel, evsel_list->threads, false);
 229 }
 230
 231 /*
 232  * Does the counter have nsecs as a unit?
 233  */
 234 static inline int nsec_counter(struct perf_evsel *evsel)
 235 {
 236         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 237             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 238                 return 1;
 239
 240         return 0;
 241 }
 242
 243 /*
 244  * Update various tracking values we maintain to print
 245  * more semantic information such as miss/hit ratios,
 246  * instruction rates, etc:
 247  */
 248 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
 249 {
 250         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
 251                 update_stats(&runtime_nsecs_stats[0], count[0]);
 252         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
 253                 update_stats(&runtime_cycles_stats[0], count[0]);
 254         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
 255                 update_stats(&runtime_stalled_cycles_stats[0], count[0]);
 256         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
 257                 update_stats(&runtime_branches_stats[0], count[0]);
 258         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
 259                 update_stats(&runtime_cacherefs_stats[0], count[0]);
 260         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
 261                 update_stats(&runtime_l1_dcache_stats[0], count[0]);
 262 }
 263
 264 /*
 265  * Read out the results of a single counter:
 266  * aggregate counts across CPUs in system-wide mode
 267  */
 268 static int read_counter_aggr(struct perf_evsel *counter)
 269 {
 270         struct perf_stat *ps = counter->priv;
 271         u64 *count = counter->counts->aggr.values;
 272         int i;
 273
 274         if (__perf_evsel__read(counter, evsel_list->cpus->nr,
 275                                evsel_list->threads->nr, scale) < 0)
 276                 return -1;
 277
 278         for (i = 0; i < 3; i++)
 279                 update_stats(&ps->res_stats[i], count[i]);
 280
 281         if (verbose) {
 282                 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 283                         event_name(counter), count[0], count[1], count[2]);
 284         }
 285
 286         /*
 287          * Save the full runtime - to allow normalization during printout:
 288          */
 289         update_shadow_stats(counter, count);
 290
 291         return 0;
 292 }
 293
 294 /*
 295  * Read out the results of a single counter:
 296  * do not aggregate counts across CPUs in system-wide mode
 297  */
 298 static int read_counter(struct perf_evsel *counter)
 299 {
 300         u64 *count;
 301         int cpu;
 302
 303         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 304                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
 305                         return -1;
 306
 307                 count = counter->counts->cpu[cpu].values;
 308
 309                 update_shadow_stats(counter, count);
 310         }
 311
 312         return 0;
 313 }
 314
 315 static int run_perf_stat(int argc __used, const char **argv)
 316 {
 317         unsigned long long t0, t1;
 318         struct perf_evsel *counter;
 319         int status = 0;
 320         int child_ready_pipe[2], go_pipe[2];
 321         const bool forks = (argc > 0);
 322         char buf;
 323
 324         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 325                 perror("failed to create pipes");
 326                 exit(1);
 327         }
 328
 329         if (forks) {
 330                 if ((child_pid = fork()) < 0)
 331                         perror("failed to fork");
 332
 333                 if (!child_pid) {
 334                         close(child_ready_pipe[0]);
 335                         close(go_pipe[1]);
 336                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 337
 338                         /*
 339                          * Do a dummy execvp to get the PLT entry resolved,
 340                          * so we avoid the resolver overhead on the real
 341                          * execvp call.
 342                          */
 343                         execvp("", (char **)argv);
 344
 345                         /*
 346                          * Tell the parent we're ready to go
 347                          */
 348                         close(child_ready_pipe[1]);
 349
 350                         /*
 351                          * Wait until the parent tells us to go.
 352                          */
 353                         if (read(go_pipe[0], &buf, 1) == -1)
 354                                 perror("unable to read pipe");
 355
 356                         execvp(argv[0], (char **)argv);
 357
 358                         perror(argv[0]);
 359                         exit(-1);
 360                 }
 361
 362                 if (target_tid == -1 && target_pid == -1 && !system_wide)
 363                         evsel_list->threads->map[0] = child_pid;
 364
 365                 /*
 366                  * Wait for the child to be ready to exec.
 367                  */
 368                 close(child_ready_pipe[1]);
 369                 close(go_pipe[0]);
 370                 if (read(child_ready_pipe[0], &buf, 1) == -1)
 371                         perror("unable to read pipe");
 372                 close(child_ready_pipe[0]);
 373         }
 374
 375         list_for_each_entry(counter, &evsel_list->entries, node) {
 376                 if (create_perf_stat_counter(counter) < 0) {
 377                         if (errno == EINVAL || errno == ENOSYS)
 378                                 continue;
 379
 380                         if (errno == EPERM || errno == EACCES) {
 381                                 error("You may not have permission to collect %sstats.\n"
 382                                       "\t Consider tweaking"
 383                                       " /proc/sys/kernel/perf_event_paranoid or running as root.",
 384                                       system_wide ? "system-wide " : "");
 385                         } else if (errno == ENOENT) {
 386                                 error("%s event is not supported. ", event_name(counter));
 387                         } else {
 388                                 error("open_counter returned with %d (%s). "
 389                                       "/bin/dmesg may provide additional information.\n",
 390                                        errno, strerror(errno));
 391                         }
 392                         if (child_pid != -1)
 393                                 kill(child_pid, SIGTERM);
 394                         die("Not all events could be opened.\n");
 395                         return -1;
 396                 }
 397         }
 398
 399         if (perf_evlist__set_filters(evsel_list)) {
 400                 error("failed to set filter with %d (%s)\n", errno,
 401                         strerror(errno));
 402                 return -1;
 403         }
 404
 405         /*
 406          * Enable counters and exec the command:
 407          */
 408         t0 = rdclock();
 409
 410         if (forks) {
 411                 close(go_pipe[1]);
 412                 wait(&status);
 413         } else {
 414                 while(!done) sleep(1);
 415         }
 416
 417         t1 = rdclock();
 418
 419         update_stats(&walltime_nsecs_stats, t1 - t0);
 420
 421         if (no_aggr) {
 422                 list_for_each_entry(counter, &evsel_list->entries, node) {
 423                         read_counter(counter);
 424                         perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
 425                 }
 426         } else {
 427                 list_for_each_entry(counter, &evsel_list->entries, node) {
 428                         read_counter_aggr(counter);
 429                         perf_evsel__close_fd(counter, evsel_list->cpus->nr,
 430                                              evsel_list->threads->nr);
 431                 }
 432         }
 433
 434         return WEXITSTATUS(status);
 435 }
 436
 437 static void print_noise_pct(double total, double avg)
 438 {
 439         double pct = 0.0;
 440
 441         if (avg)
 442                 pct = 100.0*total/avg;
 443
 444         fprintf(stderr, "  ( +-%6.2f%% )", pct);
 445 }
 446
 447 static void print_noise(struct perf_evsel *evsel, double avg)
 448 {
 449         struct perf_stat *ps;
 450
 451         if (run_count == 1)
 452                 return;
 453
 454         ps = evsel->priv;
 455         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 456 }
 457
 458 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 459 {
 460         double msecs = avg / 1e6;
 461         char cpustr[16] = { '\0', };
 462         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
 463
 464         if (no_aggr)
 465                 sprintf(cpustr, "CPU%*d%s",
 466                         csv_output ? 0 : -4,
 467                         evsel_list->cpus->map[cpu], csv_sep);
 468
 469         fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
 470
 471         if (evsel->cgrp)
 472                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 473
 474         if (csv_output)
 475                 return;
 476
 477         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 478                 fprintf(stderr, " # %8.3f CPUs utilized          ", avg / avg_stats(&walltime_nsecs_stats));
 479 }
 480
 481 static void print_stalled_cycles(int cpu, struct perf_evsel *evsel __used, double avg)
 482 {
 483         double total, ratio = 0.0;
 484         const char *color;
 485
 486         total = avg_stats(&runtime_cycles_stats[cpu]);
 487
 488         if (total)
 489                 ratio = avg / total * 100.0;
 490
 491         color = PERF_COLOR_NORMAL;
 492         if (ratio > 75.0)
 493                 color = PERF_COLOR_RED;
 494         else if (ratio > 50.0)
 495                 color = PERF_COLOR_MAGENTA;
 496         else if (ratio > 25.0)
 497                 color = PERF_COLOR_YELLOW;
 498
 499         fprintf(stderr, " #   ");
 500         color_fprintf(stderr, color, "%5.2f%%", ratio);
 501         fprintf(stderr, " of all cycles are idle ");
 502 }
 503
 504 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 505 {
 506         double total, ratio = 0.0;
 507         const char *color;
 508
 509         total = avg_stats(&runtime_branches_stats[cpu]);
 510
 511         if (total)
 512                 ratio = avg / total * 100.0;
 513
 514         color = PERF_COLOR_NORMAL;
 515         if (ratio > 20.0)
 516                 color = PERF_COLOR_RED;
 517         else if (ratio > 10.0)
 518                 color = PERF_COLOR_MAGENTA;
 519         else if (ratio > 5.0)
 520                 color = PERF_COLOR_YELLOW;
 521
 522         fprintf(stderr, " #   ");
 523         color_fprintf(stderr, color, "%5.2f%%", ratio);
 524         fprintf(stderr, " of all branches        ");
 525 }
 526
 527 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
 528 {
 529         double total, ratio = 0.0;
 530         const char *color;
 531
 532         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
 533
 534         if (total)
 535                 ratio = avg / total * 100.0;
 536
 537         color = PERF_COLOR_NORMAL;
 538         if (ratio > 20.0)
 539                 color = PERF_COLOR_RED;
 540         else if (ratio > 10.0)
 541                 color = PERF_COLOR_MAGENTA;
 542         else if (ratio > 5.0)
 543                 color = PERF_COLOR_YELLOW;
 544
 545         fprintf(stderr, " #   ");
 546         color_fprintf(stderr, color, "%5.2f%%", ratio);
 547         fprintf(stderr, " of all L1-dcache hits  ");
 548 }
 549
 550 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 551 {
 552         double total, ratio = 0.0;
 553         char cpustr[16] = { '\0', };
 554         const char *fmt;
 555
 556         if (csv_output)
 557                 fmt = "%s%.0f%s%s";
 558         else if (big_num)
 559                 fmt = "%s%'18.0f%s%-24s";
 560         else
 561                 fmt = "%s%18.0f%s%-24s";
 562
 563         if (no_aggr)
 564                 sprintf(cpustr, "CPU%*d%s",
 565                         csv_output ? 0 : -4,
 566                         evsel_list->cpus->map[cpu], csv_sep);
 567         else
 568                 cpu = 0;
 569
 570         fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
 571
 572         if (evsel->cgrp)
 573                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
 574
 575         if (csv_output)
 576                 return;
 577
 578         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 579                 total = avg_stats(&runtime_cycles_stats[cpu]);
 580
 581                 if (total)
 582                         ratio = avg / total;
 583
 584                 fprintf(stderr, " #    %4.2f  insns per cycle        ", ratio);
 585
 586                 total = avg_stats(&runtime_stalled_cycles_stats[cpu]);
 587
 588                 if (total && avg) {
 589                         ratio = total / avg;
 590                         fprintf(stderr, "\n                                            #    %4.2f  stalled cycles per insn", ratio);
 591                 }
 592
 593         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
 594                         runtime_branches_stats[cpu].n != 0) {
 595                 print_branch_misses(cpu, evsel, avg);
 596         } else if (
 597                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
 598                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
 599                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
 600                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
 601                         runtime_l1_dcache_stats[cpu].n != 0) {
 602                 print_l1_dcache_misses(cpu, evsel, avg);
 603         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
 604                         runtime_cacherefs_stats[cpu].n != 0) {
 605                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
 606
 607                 if (total)
 608                         ratio = avg * 100 / total;
 609
 610                 fprintf(stderr, " # %8.3f %% of all cache refs    ", ratio);
 611
 612         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
 613                 print_stalled_cycles(cpu, evsel, avg);
 614         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
 615                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 616
 617                 if (total)
 618                         ratio = 1.0 * avg / total;
 619
 620                 fprintf(stderr, " # %8.3f GHz                    ", ratio);
 621         } else if (runtime_nsecs_stats[cpu].n != 0) {
 622                 total = avg_stats(&runtime_nsecs_stats[cpu]);
 623
 624                 if (total)
 625                         ratio = 1000.0 * avg / total;
 626
 627                 fprintf(stderr, " # %8.3f M/sec                  ", ratio);
 628         } else {
 629                 fprintf(stderr, "                                   ");
 630         }
 631 }
 632
 633 /*
 634  * Print out the results of a single counter:
 635  * aggregated counts in system-wide mode
 636  */
 637 static void print_counter_aggr(struct perf_evsel *counter)
 638 {
 639         struct perf_stat *ps = counter->priv;
 640         double avg = avg_stats(&ps->res_stats[0]);
 641         int scaled = counter->counts->scaled;
 642
 643         if (scaled == -1) {
 644                 fprintf(stderr, "%*s%s%*s",
 645                         csv_output ? 0 : 18,
 646                         "<not counted>",
 647                         csv_sep,
 648                         csv_output ? 0 : -24,
 649                         event_name(counter));
 650
 651                 if (counter->cgrp)
 652                         fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 653
 654                 fputc('\n', stderr);
 655                 return;
 656         }
 657
 658         if (nsec_counter(counter))
 659                 nsec_printout(-1, counter, avg);
 660         else
 661                 abs_printout(-1, counter, avg);
 662
 663         if (csv_output) {
 664                 fputc('\n', stderr);
 665                 return;
 666         }
 667
 668         print_noise(counter, avg);
 669
 670         if (scaled) {
 671                 double avg_enabled, avg_running;
 672
 673                 avg_enabled = avg_stats(&ps->res_stats[1]);
 674                 avg_running = avg_stats(&ps->res_stats[2]);
 675
 676                 fprintf(stderr, "  (%.2f%%)", 100 * avg_running / avg_enabled);
 677         }
 678         fprintf(stderr, "\n");
 679 }
 680
 681 /*
 682  * Print out the results of a single counter:
 683  * does not use aggregated count in system-wide
 684  */
 685 static void print_counter(struct perf_evsel *counter)
 686 {
 687         u64 ena, run, val;
 688         int cpu;
 689
 690         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
 691                 val = counter->counts->cpu[cpu].val;
 692                 ena = counter->counts->cpu[cpu].ena;
 693                 run = counter->counts->cpu[cpu].run;
 694                 if (run == 0 || ena == 0) {
 695                         fprintf(stderr, "CPU%*d%s%*s%s%*s",
 696                                 csv_output ? 0 : -4,
 697                                 evsel_list->cpus->map[cpu], csv_sep,
 698                                 csv_output ? 0 : 18,
 699                                 "<not counted>", csv_sep,
 700                                 csv_output ? 0 : -24,
 701                                 event_name(counter));
 702
 703                         if (counter->cgrp)
 704                                 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
 705
 706                         fputc('\n', stderr);
 707                         continue;
 708                 }
 709
 710                 if (nsec_counter(counter))
 711                         nsec_printout(cpu, counter, val);
 712                 else
 713                         abs_printout(cpu, counter, val);
 714
 715                 if (!csv_output) {
 716                         print_noise(counter, 1.0);
 717
 718                         if (run != ena)
 719                                 fprintf(stderr, "  (%.2f%%)", 100.0 * run / ena);
 720                 }
 721                 fputc('\n', stderr);
 722         }
 723 }
 724
 725 static void print_stat(int argc, const char **argv)
 726 {
 727         struct perf_evsel *counter;
 728         int i;
 729
 730         fflush(stdout);
 731
 732         if (!csv_output) {
 733                 fprintf(stderr, "\n");
 734                 fprintf(stderr, " Performance counter stats for ");
 735                 if(target_pid == -1 && target_tid == -1) {
 736                         fprintf(stderr, "\'%s", argv[0]);
 737                         for (i = 1; i < argc; i++)
 738                                 fprintf(stderr, " %s", argv[i]);
 739                 } else if (target_pid != -1)
 740                         fprintf(stderr, "process id \'%d", target_pid);
 741                 else
 742                         fprintf(stderr, "thread id \'%d", target_tid);
 743
 744                 fprintf(stderr, "\'");
 745                 if (run_count > 1)
 746                         fprintf(stderr, " (%d runs)", run_count);
 747                 fprintf(stderr, ":\n\n");
 748         }
 749
 750         if (no_aggr) {
 751                 list_for_each_entry(counter, &evsel_list->entries, node)
 752                         print_counter(counter);
 753         } else {
 754                 list_for_each_entry(counter, &evsel_list->entries, node)
 755                         print_counter_aggr(counter);
 756         }
 757
 758         if (!csv_output) {
 759                 fprintf(stderr, "\n");
 760                 fprintf(stderr, " %18.9f  seconds time elapsed",
 761                                 avg_stats(&walltime_nsecs_stats)/1e9);
 762                 if (run_count > 1) {
 763                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
 764                                         avg_stats(&walltime_nsecs_stats));
 765                 }
 766                 fprintf(stderr, "\n\n");
 767         }
 768 }
 769
 770 static volatile int signr = -1;
 771
 772 static void skip_signal(int signo)
 773 {
 774         if(child_pid == -1)
 775                 done = 1;
 776
 777         signr = signo;
 778 }
 779
 780 static void sig_atexit(void)
 781 {
 782         if (child_pid != -1)
 783                 kill(child_pid, SIGTERM);
 784
 785         if (signr == -1)
 786                 return;
 787
 788         signal(signr, SIG_DFL);
 789         kill(getpid(), signr);
 790 }
 791
 792 static const char * const stat_usage[] = {
 793         "perf stat [<options>] [<command>]",
 794         NULL
 795 };
 796
 797 static int stat__set_big_num(const struct option *opt __used,
 798                              const char *s __used, int unset)
 799 {
 800         big_num_opt = unset ? 0 : 1;
 801         return 0;
 802 }
 803
 804 static const struct option options[] = {
 805         OPT_CALLBACK('e', "event", &evsel_list, "event",
 806                      "event selector. use 'perf list' to list available events",
 807                      parse_events),
 808         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
 809                      "event filter", parse_filter),
 810         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 811                     "child tasks do not inherit counters"),
 812         OPT_INTEGER('p', "pid", &target_pid,
 813                     "stat events on existing process id"),
 814         OPT_INTEGER('t', "tid", &target_tid,
 815                     "stat events on existing thread id"),
 816         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 817                     "system-wide collection from all CPUs"),
 818         OPT_BOOLEAN('c', "scale", &scale,
 819                     "scale/normalize counters"),
 820         OPT_INCR('v', "verbose", &verbose,
 821                     "be more verbose (show counter open errors, etc)"),
 822         OPT_INTEGER('r', "repeat", &run_count,
 823                     "repeat command and print average + stddev (max: 100)"),
 824         OPT_BOOLEAN('n', "null", &null_run,
 825                     "null run - dont start any counters"),
 826         OPT_BOOLEAN('d', "detailed", &detailed_run,
 827                     "detailed run - start a lot of events"),
 828         OPT_BOOLEAN('S', "sync", &sync_run,
 829                     "call sync() before starting a run"),
 830         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
 831                            "print large numbers with thousands\' separators",
 832                            stat__set_big_num),
 833         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 834                     "list of cpus to monitor in system-wide"),
 835         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
 836                     "disable CPU count aggregation"),
 837         OPT_STRING('x', "field-separator", &csv_sep, "separator",
 838                    "print counts with custom separator"),
 839         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 840                      "monitor event in cgroup name only",
 841                      parse_cgroups),
 842         OPT_END()
 843 };
 844
 845 int cmd_stat(int argc, const char **argv, const char *prefix __used)
 846 {
 847         struct perf_evsel *pos;
 848         int status = -ENOMEM;
 849
 850         setlocale(LC_ALL, "");
 851
 852         evsel_list = perf_evlist__new(NULL, NULL);
 853         if (evsel_list == NULL)
 854                 return -ENOMEM;
 855
 856         argc = parse_options(argc, argv, options, stat_usage,
 857                 PARSE_OPT_STOP_AT_NON_OPTION);
 858
 859         if (csv_sep)
 860                 csv_output = true;
 861         else
 862                 csv_sep = DEFAULT_SEPARATOR;
 863
 864         /*
 865          * let the spreadsheet do the pretty-printing
 866          */
 867         if (csv_output) {
 868                 /* User explicitely passed -B? */
 869                 if (big_num_opt == 1) {
 870                         fprintf(stderr, "-B option not supported with -x\n");
 871                         usage_with_options(stat_usage, options);
 872                 } else /* Nope, so disable big number formatting */
 873                         big_num = false;
 874         } else if (big_num_opt == 0) /* User passed --no-big-num */
 875                 big_num = false;
 876
 877         if (!argc && target_pid == -1 && target_tid == -1)
 878                 usage_with_options(stat_usage, options);
 879         if (run_count <= 0)
 880                 usage_with_options(stat_usage, options);
 881
 882         /* no_aggr, cgroup are for system-wide only */
 883         if ((no_aggr || nr_cgroups) && !system_wide) {
 884                 fprintf(stderr, "both cgroup and no-aggregation "
 885                         "modes only available in system-wide mode\n");
 886
 887                 usage_with_options(stat_usage, options);
 888         }
 889
 890         /* Set attrs and nr_counters if no event is selected and !null_run */
 891         if (detailed_run) {
 892                 size_t c;
 893
 894                 for (c = 0; c < ARRAY_SIZE(detailed_attrs); ++c) {
 895                         pos = perf_evsel__new(&detailed_attrs[c], c);
 896                         if (pos == NULL)
 897                                 goto out;
 898                         perf_evlist__add(evsel_list, pos);
 899                 }
 900         }
 901         /* Set attrs and nr_counters if no event is selected and !null_run */
 902         if (!detailed_run && !null_run && !evsel_list->nr_entries) {
 903                 size_t c;
 904
 905                 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
 906                         pos = perf_evsel__new(&default_attrs[c], c);
 907                         if (pos == NULL)
 908                                 goto out;
 909                         perf_evlist__add(evsel_list, pos);
 910                 }
 911         }
 912
 913         if (target_pid != -1)
 914                 target_tid = target_pid;
 915
 916         evsel_list->threads = thread_map__new(target_pid, target_tid);
 917         if (evsel_list->threads == NULL) {
 918                 pr_err("Problems finding threads of monitor\n");
 919                 usage_with_options(stat_usage, options);
 920         }
 921
 922         if (system_wide)
 923                 evsel_list->cpus = cpu_map__new(cpu_list);
 924         else
 925                 evsel_list->cpus = cpu_map__dummy_new();
 926
 927         if (evsel_list->cpus == NULL) {
 928                 perror("failed to parse CPUs map");
 929                 usage_with_options(stat_usage, options);
 930                 return -1;
 931         }
 932
 933         list_for_each_entry(pos, &evsel_list->entries, node) {
 934                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
 935                     perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
 936                     perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
 937                         goto out_free_fd;
 938         }
 939
 940         /*
 941          * We dont want to block the signals - that would cause
 942          * child tasks to inherit that and Ctrl-C would not work.
 943          * What we want is for Ctrl-C to work in the exec()-ed
 944          * task, but being ignored by perf stat itself:
 945          */
 946         atexit(sig_atexit);
 947         signal(SIGINT,  skip_signal);
 948         signal(SIGALRM, skip_signal);
 949         signal(SIGABRT, skip_signal);
 950
 951         status = 0;
 952         for (run_idx = 0; run_idx < run_count; run_idx++) {
 953                 if (run_count != 1 && verbose)
 954                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
 955
 956                 if (sync_run)
 957                         sync();
 958
 959                 status = run_perf_stat(argc, argv);
 960         }
 961
 962         if (status != -1)
 963                 print_stat(argc, argv);
 964 out_free_fd:
 965         list_for_each_entry(pos, &evsel_list->entries, node)
 966                 perf_evsel__free_stat_priv(pos);
 967         perf_evlist__delete_maps(evsel_list);
 968 out:
 969         perf_evlist__delete(evsel_list);
 970         return status;
 971 }