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